Re-introduce the `llama` package (#5034)

* Re-introduce the llama package

This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:

- C APIs can be called directly from Go without needing to use the previous
  "server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
  a go generate ./... step, making it easy to get up and running to hack on
  parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
  takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source

This is a big PR, but much of it is vendor code except for:

- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
  different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>

* cache: Clear old KV cache entries when evicting a slot

When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.

This change fixes two issues:
 - The KV cache fills up and runs out of space even though we think
   we are managing it correctly
 - Performance gets worse over time as we use new cache entries that
   are not hot in the processor caches

* doc: explain golang objc linker warning (#6830)

* llama: gather transitive dependencies for rocm for dist packaging (#6848)

* Refine go server makefiles to be more DRY (#6924)

This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles.  This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.

When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.

* llama: don't create extraneous directories (#6988)

* llama: Exercise the new build in CI (#6989)

Wire up some basic sanity testing in CI for the Go runner.  GPU runners are not covered yet.

* llama: Refine developer docs for Go server (#6842)

This enhances the documentation for development focusing on the new Go
server.  After we complete the transition further doc refinements
can remove the "transition" discussion.

* runner.go: Allocate batches for all sequences during init

We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.

* llama.go: Don't return nil from Tokenize on zero length input

Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.

* runner.go: Remove stop tokens from cache

If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.

However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.

This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.

By trimming the cache to the tokens that we actually return this
issue can be avoided.

* runner.go: Simplify flushing of pending tokens

* runner.go: Update TODOs

* runner.go: Don't panic when processing sequences

If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.

Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>

* runner.go: More accurately capture timings

Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.

* runner.go: Support for vision models

In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
 - Cache prompting works with images, avoiding the need to re-decode
   embeddings for every message in a conversation
 - Parallelism is supported, avoiding the need to restrict to one
   sequence at a time. (Though for now Ollama will not schedule
   them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>

* runner.go: Move Unicode checking code and add tests

* runner.go: Export external cache members

Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.

* runner.go: Image embedding cache

Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.

This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.

* llama: catch up on patches

Carry forward solar-pro and cli-unicode patches

* runner.go: Don't re-allocate memory for every batch

We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.

This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.

* runner.go: Default to classic input cache policy

The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.

However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).

This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.

For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.

* runner.go: Increase size of response channel

Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.

* llama: Add CI to verify all vendored changes have patches (#7066)

Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.

* llama: adjust clip patch for mingw utf-16 (#7065)

* llama: adjust clip patch for mingw utf-16

* llama: ensure static linking of runtime libs

Avoid runtime dependencies on non-standard libraries

* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)

These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.

* llm: Don't add BOS/EOS for tokenize requests

This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.

* runner.go: Don't cache prompts for embeddings

Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.

Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.

* runner.go: Adjust debug log levels

Add system info printed at startup and quiet down noisier logging.

* llama: fix compiler flag differences (#7082)

Adjust the flags for the new Go server to more closely match the
generate flow

* llama: refine developer docs (#7121)

* llama: doc and example clean up (#7122)

* llama: doc and example clean up

* llama: Move new dockerfile into llama dir

Temporary home until we fully transition to the Go server

* llama: runner doc cleanup

* llama.go: Add description for Tokenize error case

---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>

llama/ggml-cuda/fattn-tile-f32.cuh

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/fattn-vec-f16.cuh

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "common.cuh"
+#include "fattn-common.cuh"
+
+template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_vec_ext_f16(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const float logit_softcap,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int nb21,
+        const int nb22,
+        const int nb23,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#ifdef FP16_AVAILABLE
+    // Skip unused kernel variants for faster compilation:
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    constexpr vec_dot_KQ_f16_t vec_dot_KQ = get_vec_dot_KQ_f16<D>(type_K);
+    constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
+    constexpr dequantize_1_f16_t dequantize_1_v = get_dequantize_1_f16(type_V);
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    Q += nb02* blockIdx.y              + nb01*ic0;
+    K += nb12*(blockIdx.y / gqa_ratio);
+    V += nb22*(blockIdx.y / gqa_ratio);
+
+    const half * maskh = (const half   *)  mask + ne11*ic0;
+
+    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const half  slopeh = __float2half(slopef);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+    constexpr int nwarps = D / WARP_SIZE;
+    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ half KQ[ncols*D];
+    half2 * KQ2 = (half2 *) KQ;
+
+    half kqmax[ncols];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqmax[j] = -HALF_MAX_HALF;
+    }
+    half kqsum[ncols] = {0.0f};
+
+    __shared__ half kqmax_shared[ncols][WARP_SIZE];
+    __shared__ half kqsum_shared[ncols][WARP_SIZE];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        if (threadIdx.y == 0) {
+            kqmax_shared[j][threadIdx.x] = -HALF_MAX_HALF;
+            kqsum_shared[j][threadIdx.x] = 0.0f;
+        }
+    }
+    __syncthreads();
+
+    // Convert Q to half2 (f16 K) or q8_1 (quantized K) and store in registers:
+    half2  Q_h2[ncols][D/(2*WARP_SIZE)];
+    int   Q_i32[ncols][D/(sizeof(int)*QK8_1) == 0 ? 1 : D/(sizeof(int)*QK8_1)];
+    half2  Q_ds[ncols][D/QK8_1 == 0 ? 1 : D/QK8_1];
+    if (Q_q8_1) {
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            if (j0 + nwarps > ncols && j >= ncols) {
+                break;
+            }
+
+            // Reuse KQ as temporary storage for converting Q to q8_1:
+            int   * tmp_q_i32 = (int   *) &KQ[j*D];
+            half2 * tmp_q_ds  = (half2 *) (tmp_q_i32 + D/sizeof(int));
+
+            // Set memory to zero if out of bounds:
+            if (ncols > 2 && ic0 + j >= ne01) {
+#pragma unroll
+                for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
+                    const int i = i0 + threadIdx.x;
+
+                    tmp_q_i32[i] = 0;
+                }
+                if (threadIdx.x < D/QK8_1) {
+                    tmp_q_ds[threadIdx.x] = make_half2(0.0f, 0.0f);
+                }
+                continue;
+            }
+
+            const float * Q_f = (const float *) (Q + j*nb01);
+#pragma unroll
+            for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
+                quantize_q8_1_to_shared<half2>(Q_f + 4*i0, scale, tmp_q_i32, tmp_q_ds);
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            int   * tmp_q_i32 = (int   *) &KQ[j*D];
+            half2 * tmp_q_ds  = (half2 *) (tmp_q_i32 + D/sizeof(int));
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                Q_i32[j][i0/WARP_SIZE] = tmp_q_i32[i];
+                Q_ds[j][i0/WARP_SIZE]  = tmp_q_ds[i/QI8_1];
+            }
+        }
+
+        __syncthreads();
+    } else {
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            const float2 * Q_f2_j = (const float2 *) (Q + j*nb01);
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const float2 tmp = ncols <= 2 || ic0 + j < ne01 ? Q_f2_j[i] : make_float2(0.0f, 0.0f);
+                Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
+            }
+        }
+    }
+
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        KQ[j*D + tid] = -HALF_MAX_HALF;
+    }
+
+    half2 VKQ[ncols] = {{0.0f, 0.0f}};
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
+        // see https://github.com/ggerganov/llama.cpp/pull/7061 .
+        // Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
+        half kqmax_new = kqmax[0];
+        half kqmax_new_arr[ncols];
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            kqmax_new_arr[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
+                break;
+            }
+
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                half sum = vec_dot_KQ(K + (k_VKQ_0 + i_KQ)*nb11, Q_h2[j], Q_i32[j], Q_ds[j]);
+                sum = warp_reduce_sum(sum);
+
+                if (use_logit_softcap) {
+                    sum = logit_softcap*tanhf(sum);
+                }
+
+                sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
+
+                if (ncols == 1) {
+                    kqmax_new        = ggml_cuda_hmax(kqmax_new,        sum);
+                } else {
+                    kqmax_new_arr[j] = ggml_cuda_hmax(kqmax_new_arr[j], sum);
+                }
+
+                if (threadIdx.x == 0) {
+                    KQ[j*D + i_KQ] = sum;
+                }
+            }
+        }
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j];
+
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+            if (threadIdx.x == 0) {
+                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            half kqmax_new_j = kqmax_shared[j][threadIdx.x];
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
+            kqmax[j] = kqmax_new_j;
+
+            const half val = hexp(KQ[j*D + tid] - kqmax[j]);
+            kqsum[j] = kqsum[j]*KQ_max_scale + val;
+            KQ[j*D + tid] = val;
+
+            VKQ[j] *= __half2half2(KQ_max_scale);
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < D; k0 += 2) {
+            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) {
+                break;
+            }
+
+            half2 V_k;
+            reinterpret_cast<half&>(V_k.x) = dequantize_1_v(V + (k_VKQ_0 + k0 + 0)*nb21, tid);
+            reinterpret_cast<half&>(V_k.y) = dequantize_1_v(V + (k_VKQ_0 + k0 + 1)*nb21, tid);
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                VKQ[j] += V_k*KQ2[j*(D/2) + k0/2];
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqsum[j] = warp_reduce_sum(kqsum[j]);
+        if (threadIdx.x == 0) {
+            kqsum_shared[j][threadIdx.y] = kqsum[j];
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        if (ncols > 2 && ic0 + j_VKQ >= ne01) {
+            break;
+        }
+
+        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
+        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
+
+        half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
+        if (parallel_blocks == 1) {
+            dst_val /= kqsum[j_VKQ];
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+    }
+
+    if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
+        dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
+    }
+#else
+   NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+}
+
+template <int D, int cols_per_block, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
+void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    constexpr int nwarps = D/WARP_SIZE;
+    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>;
+    constexpr bool need_f16_K = D != 128;
+    constexpr bool need_f16_V = D != 128 && D != 64;
+    launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, need_f16_K, need_f16_V);
+}
+
+template <int D, ggml_type type_K, ggml_type type_V>
+void ggml_cuda_flash_attn_ext_vec_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+    const ggml_tensor * K   = dst->src[1];
+    const ggml_tensor * V   = dst->src[2];
+
+    const int32_t precision = KQV->op_params[3];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    GGML_ASSERT(K->type == type_K);
+    GGML_ASSERT(V->type == type_V);
+
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
+    if (Q->ne[1] == 1) {
+        constexpr int cols_per_block  = 1;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] == 2) {
+        constexpr int cols_per_block  = 2;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 4) {
+        constexpr int cols_per_block  = 4;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 8) {
+        constexpr int cols_per_block  = 8;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    constexpr int cols_per_block  = 8;
+    constexpr int parallel_blocks = 1;
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+    } else {
+        constexpr bool use_logit_softcap = true;
+        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+    }
+}
+
+#define DECL_FATTN_VEC_F16_CASE(D, type_K, type_V)                          \
+    template void ggml_cuda_flash_attn_ext_vec_f16_case                     \
+    <D, type_K, type_V>(ggml_backend_cuda_context & ctx, ggml_tensor * dst) \
+
+extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16);
+
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_0);
+
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_1);
+
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_0);
+
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_1);
+
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q8_0);
+
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_F16);
+
+extern DECL_FATTN_VEC_F16_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16);

llama/ggml-cuda/fattn-vec-f32.cuh

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "common.cuh"
+#include "fattn-common.cuh"
+
+template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_vec_ext_f32(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const float logit_softcap,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int nb21,
+        const int nb22,
+        const int nb23,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+    // Skip unused kernel variants for faster compilation:
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    constexpr vec_dot_KQ_f32_t vec_dot_KQ = get_vec_dot_KQ_f32<D>(type_K);
+    constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
+    constexpr dequantize_1_f32_t dequantize_1_v = get_dequantize_1_f32(type_V);
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    Q += nb02* blockIdx.y              + nb01*ic0;
+    K += nb12*(blockIdx.y / gqa_ratio);
+    V += nb22*(blockIdx.y / gqa_ratio); // K and V have same shape
+    const half * maskh = (const half   *)  mask + ne11*ic0;
+
+    const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+    constexpr int nwarps = D / WARP_SIZE;
+    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ float KQ[ncols*D];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        KQ[j*D + tid] = -FLT_MAX/2.0f;
+    }
+
+    float kqmax[ncols];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqmax[j] = -FLT_MAX/2.0f;
+    }
+    float kqsum[ncols] = {0.0f};
+
+    __shared__ float kqmax_shared[ncols][WARP_SIZE];
+    __shared__ float kqsum_shared[ncols][WARP_SIZE];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        if (threadIdx.y == 0) {
+            kqmax_shared[j][threadIdx.x] = -FLT_MAX/2.0f;
+            kqsum_shared[j][threadIdx.x] = 0.0f;
+        }
+    }
+    __syncthreads();
+
+    // Convert Q to float2 (f16 K) or q8_1 (quantized K) and store in registers:
+    float2  Q_f2[ncols][D/(2*WARP_SIZE)];
+    int    Q_i32[ncols][D/(sizeof(int)*QK8_1) == 0 ? 1 : D >= D/(sizeof(int)*QK8_1)];
+    float2  Q_ds[ncols][D/QK8_1 == 0 ? 1 : D/QK8_1];
+    if (Q_q8_1) {
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            if (j0 + nwarps > ncols && j >= ncols) {
+                break;
+            }
+
+            // Reuse KQ as temporary storage for converting Q to q8_1:
+            int    * tmp_q_i32 = (int    *) &KQ[j*D];
+            float2 * tmp_q_ds  = (float2 *) (tmp_q_i32 + D/sizeof(int));
+
+            // Set memory to zero if out of bounds:
+            if (ncols > 2 && ic0 + j >= ne01) {
+#pragma unroll
+                for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
+                    const int i = i0 + threadIdx.x;
+
+                    tmp_q_i32[i] = 0;
+                }
+                if (threadIdx.x < D/QK8_1) {
+                    tmp_q_ds[threadIdx.x] = make_float2(0.0f, 0.0f);
+                }
+                continue;
+            }
+
+            const float * Q_f = (const float *) (Q + j*nb01);
+#pragma unroll
+            for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
+                quantize_q8_1_to_shared<float2>(Q_f + 4*i0, scale, tmp_q_i32, tmp_q_ds);
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            int    * tmp_q_i32 = (int    *) &KQ[j*D];
+            float2 * tmp_q_ds  = (float2 *) (tmp_q_i32 + D/sizeof(int));
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                Q_i32[j][i0/WARP_SIZE] = tmp_q_i32[i];
+                Q_ds[j][i0/WARP_SIZE]  = tmp_q_ds[i/QI8_1];
+            }
+        }
+
+        __syncthreads();
+    } else {
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            const float2 * Q_f2_j = (const float2 *) (Q + j*nb01);
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                Q_f2[j][i0/WARP_SIZE]    = ncols <= 2 || ic0 + j < ne01 ? Q_f2_j[i] : make_float2(0.0f, 0.0f);
+                Q_f2[j][i0/WARP_SIZE].x *= scale;
+                Q_f2[j][i0/WARP_SIZE].y *= scale;
+            }
+        }
+    }
+
+    float VKQ[ncols] = {0.0f};
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        float kqmax_new_arr[ncols];
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            kqmax_new_arr[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
+                break;
+            }
+
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                float sum = vec_dot_KQ(K + (k_VKQ_0 + i_KQ)*nb11, Q_f2[j], Q_i32[j], Q_ds[j]);
+                sum = warp_reduce_sum(sum);
+
+                if (use_logit_softcap) {
+                    sum = logit_softcap*tanhf(sum);
+                }
+
+                sum += mask ? slope*__half2float(maskh[j*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
+
+                kqmax_new_arr[j] = fmaxf(kqmax_new_arr[j], sum);
+
+                if (threadIdx.x == 0) {
+                    KQ[j*D + i_KQ] = sum;
+                }
+            }
+        }
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            float kqmax_new_j = kqmax_new_arr[j];
+
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+            if (threadIdx.x == 0) {
+                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            float kqmax_new_j = kqmax_shared[j][threadIdx.x];
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const float KQ_max_scale = expf(kqmax[j] - kqmax_new_j);
+            kqmax[j] = kqmax_new_j;
+
+            const float val = expf(KQ[j*D + tid] - kqmax[j]);
+            kqsum[j] = kqsum[j]*KQ_max_scale + val;
+            KQ[j*D + tid] = val;
+
+            VKQ[j] *= KQ_max_scale;
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k = 0; k < D; ++k) {
+            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k >= ne11) {
+                break;
+            }
+
+            const float V_ki = dequantize_1_v(V + (k_VKQ_0 + k)*nb21, tid);
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                VKQ[j] += V_ki*KQ[j*D + k];
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqsum[j] = warp_reduce_sum(kqsum[j]);
+        if (threadIdx.x == 0) {
+            kqsum_shared[j][threadIdx.y] = kqsum[j];
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        if (ncols > 2 && ic0 + j_VKQ >= ne01) {
+            break;
+        }
+
+        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
+        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
+
+        float dst_val = VKQ[j_VKQ];
+        if (parallel_blocks == 1) {
+            dst_val /= kqsum[j_VKQ];
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+    }
+
+    if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
+        dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
+    }
+}
+
+template <int D, int cols_per_block, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
+void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    constexpr int nwarps = D/WARP_SIZE;
+    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>;
+    constexpr bool need_f16_K = D != 128;
+    constexpr bool need_f16_V = D != 128 && D != 64;
+    launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, need_f16_K, need_f16_V);
+}
+
+template <int D, ggml_type type_K, ggml_type type_V>
+void ggml_cuda_flash_attn_ext_vec_f32_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+    const ggml_tensor * K   = dst->src[1];
+    const ggml_tensor * V   = dst->src[2];
+
+    GGML_ASSERT(K->type == type_K);
+    GGML_ASSERT(V->type == type_V);
+
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
+    if (Q->ne[1] == 1) {
+        constexpr int cols_per_block  = 1;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] == 2) {
+        constexpr int cols_per_block  = 2;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 4) {
+        constexpr int cols_per_block  = 4;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 8) {
+        constexpr int cols_per_block  = 8;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    constexpr int cols_per_block  = 8;
+    constexpr int parallel_blocks = 1;
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+    } else {
+        constexpr bool use_logit_softcap = true;
+        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+    }
+}
+
+#define DECL_FATTN_VEC_F32_CASE(D, type_K, type_V)                          \
+    template void ggml_cuda_flash_attn_ext_vec_f32_case                     \
+    <D, type_K, type_V>(ggml_backend_cuda_context & ctx, ggml_tensor * dst) \
+
+extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16);
+
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_0);
+
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_1);
+
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_0);
+
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_1);
+
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q8_0);
+
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_F16);
+
+extern DECL_FATTN_VEC_F32_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16);

llama/ggml-cuda/fattn-wmma-f16.cuh

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "common.cuh"
+#include "fattn-common.cuh"
+
+#ifdef FP16_MMA_AVAILABLE
+#include <mma.h>
+#endif // FP16_MMA_AVAILABLE
+
+// D == head size, VKQ_stride == num VKQ rows calculated in parallel:
+template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t, bool use_logit_softcap>
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(nwarps*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_ext_f16(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const float logit_softcap,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int nb21,
+        const int nb22,
+        const int nb23,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#ifdef FP16_MMA_AVAILABLE
+    // Skip unused kernel variants for faster compilation:
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = ncols*(blockIdx.x / parallel_blocks); // Index of the first Q/QKV column to work on.
+    const int ip  =        blockIdx.x % parallel_blocks;  // Index in group of blocks running for the same column in parallel.
+
+    static_assert(D <= FATTN_KQ_STRIDE, "D must be <= FATTN_KQ_STRIDE.");
+    static_assert(ncols == 8 || ncols % 16 == 0, "ncols must be 8 or a multiple of 16.");
+    constexpr int frag_m = ncols == 8 ? 32 : 16;
+    constexpr int frag_n = ncols == 8 ?  8 : 16;
+    static_assert(D % frag_m == 0, "If ncols == 8 then D % frag_m must be 0.");
+    typedef nvcuda::wmma::fragment<nvcuda::wmma::matrix_a,    frag_m, frag_n, 16, half, nvcuda::wmma::row_major> frag_a_K;
+    typedef nvcuda::wmma::fragment<nvcuda::wmma::matrix_a,    frag_m, frag_n, 16, half, nvcuda::wmma::col_major> frag_a_V;
+    typedef nvcuda::wmma::fragment<nvcuda::wmma::matrix_b,    frag_m, frag_n, 16, half, nvcuda::wmma::col_major> frag_b;
+    typedef nvcuda::wmma::fragment<nvcuda::wmma::accumulator, frag_m, frag_n, 16, KQ_acc_t>                      frag_c_KQ;
+    typedef nvcuda::wmma::fragment<nvcuda::wmma::accumulator, frag_m, frag_n, 16, half>                          frag_c_VKQ;
+
+    constexpr int KQ_stride_tc  = nwarps*frag_m; // Number of KQ rows calculated in parallel.
+    constexpr int VKQ_ratio = KQ_stride_tc/VKQ_stride; // Number of parallel VKQ accumulators needed to keep all warps busy.
+    static_assert(VKQ_ratio <= nwarps, "VKQ_ratio must be <= nwarps.");
+
+    // Pad internal representation of KQ, KQV to reduce shared memory bank conflicts:
+    constexpr int D_padded = D + 8;
+    constexpr int kqs_padded = FATTN_KQ_STRIDE + 8;
+    constexpr int kqar = sizeof(KQ_acc_t)/sizeof(half);
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float * Q_f   = (const float *) (Q + nb02* blockIdx.y              + nb01*ic0);
+    const half  * K_h   = (const half  *) (K + nb12*(blockIdx.y / gqa_ratio));
+    const half  * V_h   = (const half  *) (V + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half  * maskh = (const half  *)  mask + (nb31/sizeof(half))* ic0;
+    const half2 * mask2 = (const half2 *)  mask + (nb31/sizeof(half))*(ic0/2);
+
+    const int stride_Q  = nb01 / sizeof(float);
+    const int stride_KV = nb11 / sizeof(half);
+
+    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const half  slopeh = __float2half(slopef);
+    const half2 slope2 = make_half2(slopef, slopef);
+
+    const half2 logit_softcap_2 = make_half2(logit_softcap, logit_softcap);
+
+    frag_b Q_b[D/16][ncols/frag_n];
+
+    // A single buffer for temporarily holding tiles of KQ and VKQ parts:
+    constexpr int mem_KQ = ncols*kqs_padded*kqar;
+    constexpr int mem_VKQ_parts = VKQ_ratio*ncols*D_padded;
+    __shared__ half KQ[mem_KQ >= mem_VKQ_parts ? mem_KQ : mem_VKQ_parts];
+    float * KQ_f = (float *) KQ;
+    half2 * KQ2 = (half2 *) KQ;
+
+    float    KQ_rowsum_f[ncols/nwarps] = {0.0f};
+    float       KQ_max_f[ncols/nwarps];
+    float KQ_max_scale_f[ncols/nwarps] = {0.0f};
+
+#pragma unroll
+    for (int j = 0; j < ncols/nwarps; ++j) {
+        KQ_max_f[j] = -FLT_MAX/2.0f;
+    }
+
+    half2    KQ_rowsum_h2[ncols/nwarps] = {{0.0f, 0.0f}};
+    half2       KQ_max_h2[ncols/nwarps];
+    half2 KQ_max_scale_h2[ncols/nwarps] = {{0.0f, 0.0f}};
+
+#pragma unroll
+    for (int j = 0; j < ncols/nwarps; ++j) {
+        KQ_max_h2[j] = make_half2(-HALF_MAX_HALF, -HALF_MAX_HALF);
+    }
+
+    __shared__ half VKQ[ncols*D_padded]; // Accumulator for final VKQ slice.
+    half2 * VKQ2 = (half2 *) VKQ;
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+            if (i0 + WARP_SIZE > D/2 && i >= D/2) {
+                break;
+            }
+            VKQ2[j*(D_padded/2) + i] = make_half2(0.0f, 0.0f);
+        }
+    }
+
+    // Convert Q to half and apply scale, temporarily store in KQ:
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+#pragma unroll
+        for (int i0 = 0; i0 < D; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+            if (i0 + WARP_SIZE > D && i >= D) {
+                break;
+            }
+            KQ[j*D_padded + i] = ic0 + j < ne01 ? Q_f[j*stride_Q + i] * scale : 0.0f;
+        }
+    }
+
+    __syncthreads();
+
+    // Load Q into tensor core fragments/registers since it will be used frequently:
+#pragma unroll
+    for (int i0 = 0; i0 < D; i0 += 16) {
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += frag_n) {
+            nvcuda::wmma::load_matrix_sync(Q_b[i0/16][j0/frag_n], KQ + j0*D_padded + i0, D_padded);
+        }
+    }
+
+    __syncthreads();
+
+    // Iterate over ne11 == previous tokens:
+    for (int k_VKQ_0 = ip*FATTN_KQ_STRIDE; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE) {
+        // Calculate tile of KQ:
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE; i_KQ_0 += KQ_stride_tc) {
+            frag_c_KQ KQ_c[ncols/frag_n];
+#pragma unroll
+            for (int j = 0; j < ncols/frag_n; ++j) {
+                nvcuda::wmma::fill_fragment(KQ_c[j], 0.0f);
+            }
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 16) {
+                frag_a_K K_a;
+                nvcuda::wmma::load_matrix_sync(K_a, K_h + (k_VKQ_0 + i_KQ_0 + frag_m*threadIdx.y)*stride_KV + k_KQ_0, stride_KV);
+#pragma unroll
+                for (int j = 0; j < ncols/frag_n; ++j) {
+                    nvcuda::wmma::mma_sync(KQ_c[j], K_a, Q_b[k_KQ_0/16][j], KQ_c[j]);
+                }
+            }
+#pragma unroll
+            for (int j0 = 0; j0 < ncols; j0 += frag_n) {
+                nvcuda::wmma::store_matrix_sync((KQ_acc_t *) KQ + j0*kqs_padded + i_KQ_0 + frag_m*threadIdx.y, KQ_c[j0/frag_n], kqs_padded, nvcuda::wmma::mem_col_major);
+            }
+        }
+
+        __syncthreads();
+
+        // Calculate softmax for each KQ column using the current max. value.
+        // The divisor is stored in KQ_rowsum and will be applied at the end.
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            if (std::is_same<KQ_acc_t, float>::value) {
+                float KQ_f_tmp[FATTN_KQ_STRIDE / WARP_SIZE];
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    KQ_f_tmp[k0/WARP_SIZE] = KQ_f[j*kqs_padded + k];
+
+                    if (use_logit_softcap) {
+                        KQ_f_tmp[k0/WARP_SIZE] = logit_softcap*tanhf(KQ_f_tmp[k0/WARP_SIZE]);
+                    }
+                }
+
+                float KQ_max_new = KQ_max_f[j0/nwarps];
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    KQ_f_tmp[k0/WARP_SIZE] += mask ? __half2float(slopeh*maskh[j*(nb31/sizeof(half)) + k_VKQ_0 + k]) : 0.0f;
+                    KQ_max_new = max(KQ_max_new, KQ_f_tmp[k0/WARP_SIZE]);
+                }
+                KQ_max_new = warp_reduce_max(KQ_max_new);
+
+                const float diff = KQ_max_f[j0/nwarps] - KQ_max_new;
+                KQ_max_scale_f[j0/nwarps] = expf(diff);
+                if (diff <= SOFTMAX_FTZ_THRESHOLD) {
+                    KQ_max_scale_f[j0/nwarps] = 0.0f;
+                }
+                KQ_max_f[j0/nwarps] = KQ_max_new;
+
+                float KQ_rowsum_add = 0.0f;
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    const float diff = KQ_f_tmp[k0/WARP_SIZE] - KQ_max_f[j0/nwarps];
+                    KQ_f_tmp[k0/WARP_SIZE] = expf(diff);
+                    if (diff <= SOFTMAX_FTZ_THRESHOLD) {
+                        KQ_f_tmp[k0/WARP_SIZE] = 0.0f;
+                    }
+                    KQ_rowsum_add += KQ_f_tmp[k0/WARP_SIZE];
+                    KQ[j*(kqar*kqs_padded) + k] = KQ_f_tmp[k0/WARP_SIZE];
+                }
+                KQ_rowsum_add = warp_reduce_sum(KQ_rowsum_add);
+
+                // Scale previous KQ_rowsum to account for a potential increase in KQ_max:
+                KQ_rowsum_f[j0/nwarps] = KQ_max_scale_f[j0/nwarps]*KQ_rowsum_f[j0/nwarps] + KQ_rowsum_add;
+            } else {
+                half2 KQ2_tmp[FATTN_KQ_STRIDE/(2*WARP_SIZE)];
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    KQ2_tmp[k0/WARP_SIZE] = KQ2[j*(kqs_padded/2) + k];
+
+                    if (use_logit_softcap) {
+                        // There is no dedicated tangens hyperbolicus function for half2.
+                        KQ2_tmp[k0/WARP_SIZE] = h2exp(KQ2_tmp[k0/WARP_SIZE]*make_half2(2.0f, 2.0f));
+                        KQ2_tmp[k0/WARP_SIZE] = (KQ2_tmp[k0/WARP_SIZE] - make_half2(1.0f, 1.0f))
+                                               /(KQ2_tmp[k0/WARP_SIZE] + make_half2(1.0f, 1.0f));
+
+                        KQ2_tmp[k0/WARP_SIZE] *= logit_softcap_2;
+                    }
+                }
+
+                half2 KQ_max_new = KQ_max_h2[j0/nwarps];
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    KQ2_tmp[k0/WARP_SIZE] += mask ? slope2*mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f);
+                    KQ_max_new = ggml_cuda_hmax2(KQ_max_new, KQ2_tmp[k0/WARP_SIZE]);
+                }
+                KQ_max_new = __half2half2(warp_reduce_max(ggml_cuda_hmax(__low2half(KQ_max_new), __high2half(KQ_max_new))));
+                const half2 diff = KQ_max_h2[j0/nwarps] - KQ_max_new;
+                KQ_max_scale_h2[j0/nwarps] = h2exp(diff);
+                const uint32_t ftz_mask = __hgt2_mask(diff, make_half2(SOFTMAX_FTZ_THRESHOLD, SOFTMAX_FTZ_THRESHOLD));
+                *((uint32_t *) &KQ_max_scale_h2[j0/nwarps]) &= ftz_mask;
+                KQ_max_h2[j0/nwarps] = KQ_max_new;
+
+                half2 KQ_rowsum_add = make_half2(0.0f, 0.0f);
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    const half2 diff = KQ2_tmp[k0/WARP_SIZE] - KQ_max_h2[j0/nwarps];
+                    KQ2_tmp[k0/WARP_SIZE] = h2exp(diff);
+                    const uint32_t ftz_mask = __hgt2_mask(diff, make_half2(SOFTMAX_FTZ_THRESHOLD, SOFTMAX_FTZ_THRESHOLD));
+                    *((uint32_t *) &KQ2_tmp[k0/WARP_SIZE]) &= ftz_mask;
+                    KQ_rowsum_add += KQ2_tmp[k0/WARP_SIZE];
+                    KQ2[j*(kqs_padded/2) + k] = KQ2_tmp[k0/WARP_SIZE];
+                }
+                KQ_rowsum_add = warp_reduce_sum(KQ_rowsum_add);
+
+                // Scale previous KQ_rowsum to account for a potential increase in KQ_max:
+                KQ_rowsum_h2[j0/nwarps] = KQ_max_scale_h2[j0/nwarps]*KQ_rowsum_h2[j0/nwarps] + KQ_rowsum_add;
+            }
+        }
+
+        __syncthreads();
+
+        frag_b KQ_b[FATTN_KQ_STRIDE/(VKQ_ratio*16)][ncols/frag_n];
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += frag_n) {
+#pragma unroll
+            for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += VKQ_ratio*16) {
+                const int k = k0 + (threadIdx.y % VKQ_ratio)*16;
+                nvcuda::wmma::load_matrix_sync(
+                    KQ_b[k0/(VKQ_ratio*16)][j0/frag_n],
+                    KQ + j0*(kqar*kqs_padded) + k,
+                    kqar*kqs_padded);
+            }
+        }
+
+        frag_c_VKQ VKQ_c[D/VKQ_stride][ncols/frag_n];
+#pragma unroll
+        for (int i_VKQ_0 = 0; i_VKQ_0 < D; i_VKQ_0 += VKQ_stride) {
+#pragma unroll
+            for (int j = 0; j < ncols/frag_n; ++j) {
+                nvcuda::wmma::fill_fragment(VKQ_c[i_VKQ_0/VKQ_stride][j], 0.0f);
+            }
+
+#pragma unroll
+            for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += VKQ_ratio*16) {
+                const int k = k0 + (threadIdx.y % VKQ_ratio)*16;
+
+                frag_a_V v_a;
+                nvcuda::wmma::load_matrix_sync(v_a, V_h + (k_VKQ_0 + k)*stride_KV + i_VKQ_0 + frag_m*(threadIdx.y/VKQ_ratio), stride_KV);
+#pragma unroll
+                for (int j = 0; j < ncols/frag_n; ++j) {
+                    nvcuda::wmma::mma_sync(VKQ_c[i_VKQ_0/VKQ_stride][j], v_a, KQ_b[k0/(VKQ_ratio*16)][j], VKQ_c[i_VKQ_0/VKQ_stride][j]);
+                }
+            }
+        }
+
+        __syncthreads();
+
+        const int offset_k = (threadIdx.y % VKQ_ratio) * (ncols*D_padded);
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += VKQ_stride) {
+#pragma unroll
+            for (int j0 = 0; j0 < ncols; j0 += frag_n) {
+                nvcuda::wmma::store_matrix_sync(
+                    KQ + offset_k + j0*D_padded + i_KQ_0 + frag_m*(threadIdx.y/VKQ_ratio),
+                    VKQ_c[i_KQ_0/VKQ_stride][j0/frag_n],
+                    D_padded, nvcuda::wmma::mem_col_major);
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            half2 VKQ_scale;
+            if (std::is_same<KQ_acc_t, float>::value) {
+                VKQ_scale = make_half2(KQ_max_scale_f[j0/nwarps], KQ_max_scale_f[j0/nwarps]);
+            } else {
+                VKQ_scale = KQ_max_scale_h2[j0/nwarps];
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+                if (i0 + WARP_SIZE > D/2 && i >= D/2) {
+                    break;
+                }
+
+                half2 VKQ_add = make_half2(0.0f, 0.0f);
+#pragma unroll
+                for (int l = 0; l < VKQ_ratio; ++l) {
+                    VKQ_add += KQ2[l*(ncols*D_padded/2) + j*(D_padded/2) + i];
+                }
+                VKQ2[j*(D_padded/2) + i] = VKQ_scale*VKQ2[j*(D_padded/2) + i] + VKQ_add;
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j_VKQ = j0 + threadIdx.y;
+        if (ic0 + j_VKQ >= ne01) {
+            return;
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+
+        float KQ_rowsum_j;
+        if (std::is_same<KQ_acc_t, float>::value) {
+            KQ_rowsum_j = KQ_rowsum_f[j0/nwarps];
+        } else {
+            KQ_rowsum_j = __low2float(KQ_rowsum_h2[j0/nwarps]) + __high2float(KQ_rowsum_h2[j0/nwarps]);
+        }
+
+#pragma unroll
+        for (int i0 = 0; i0 < D; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+            if (i0 + WARP_SIZE > D && i >= D) {
+                break;
+            }
+            float dst_val = VKQ[j_VKQ*D_padded + i];
+            if (parallel_blocks == 1) {
+                dst_val /= KQ_rowsum_j;
+            }
+            dst[j_dst*gridDim.y*D + blockIdx.y*D + i] = dst_val;
+        }
+
+        if (parallel_blocks == 1 || threadIdx.x != 0) {
+            continue;
+        }
+
+        float2 dst_meta_val;
+        if (std::is_same<KQ_acc_t, float>::value) {
+            dst_meta_val.x = KQ_max_f[j0/nwarps];
+        } else {
+            dst_meta_val.x = __low2float(KQ_max_h2[j0/nwarps]);
+        }
+        dst_meta_val.y = KQ_rowsum_j;
+        dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = dst_meta_val;
+    }
+#else
+   NO_DEVICE_CODE;
+#endif // FP16_MMA_AVAILABLE
+}
+
+constexpr int get_max_power_of_2(int x) {
+    return x % 2 == 0 ? 2*get_max_power_of_2(x/2) : 1;
+}
+
+static_assert(get_max_power_of_2(1) == 1, "Test failed.");
+static_assert(get_max_power_of_2(2) == 2, "Test failed.");
+static_assert(get_max_power_of_2(4) == 4, "Test failed.");
+static_assert(get_max_power_of_2(6) == 2, "Test failed.");
+
+// Number of VKQ rows calculated in parallel:
+constexpr int get_VKQ_stride(int D, int nwarps, int frag_m) {
+    return (get_max_power_of_2(D/frag_m) < nwarps ? get_max_power_of_2(D/frag_m) : nwarps)*frag_m;
+}
+
+static_assert(get_VKQ_stride(128, 1, 32) ==  32, "Test failed.");
+static_assert(get_VKQ_stride(128, 2, 32) ==  64, "Test failed.");
+static_assert(get_VKQ_stride(128, 4, 32) == 128, "Test failed.");
+static_assert(get_VKQ_stride( 64, 1, 32) ==  32, "Test failed.");
+static_assert(get_VKQ_stride( 64, 2, 32) ==  64, "Test failed.");
+static_assert(get_VKQ_stride( 64, 4, 32) ==  64, "Test failed.");
+static_assert(get_VKQ_stride( 80, 1, 16) ==  16, "Test failed.");
+static_assert(get_VKQ_stride( 80, 2, 16) ==  16, "Test failed.");
+static_assert(get_VKQ_stride( 80, 4, 16) ==  16, "Test failed.");
+
+template <int D, int cols_per_block, typename KQ_acc_t>
+void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    constexpr int nwarps = 4;
+
+    constexpr int frag_m = cols_per_block == 8 && D % 32 == 0 ? 32 : 16;
+    const int blocks_num_pb1 = ((Q->ne[1] + cols_per_block - 1) / cols_per_block)*Q->ne[2]*Q->ne[3];
+    const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm;
+
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
+    if (4*blocks_num_pb1 < 2*nsm) {
+        constexpr int parallel_blocks = 4;
+        fattn_kernel_t fattn_kernel;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            fattn_kernel = flash_attn_ext_f16<
+                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+        } else {
+            constexpr bool use_logit_softcap = true;
+            fattn_kernel = flash_attn_ext_f16<
+                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+        }
+        launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
+        return;
+    }
+    if (2*blocks_num_pb1 < 2*nsm) {
+        constexpr int parallel_blocks = 2;
+        fattn_kernel_t fattn_kernel;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            fattn_kernel = flash_attn_ext_f16<
+                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+        } else {
+            constexpr bool use_logit_softcap = true;
+            fattn_kernel = flash_attn_ext_f16<
+                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+        }
+        launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
+        return;
+    }
+    constexpr int parallel_blocks = 1;
+    fattn_kernel_t fattn_kernel;
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        fattn_kernel = flash_attn_ext_f16<
+            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+    } else {
+        constexpr bool use_logit_softcap = true;
+        fattn_kernel = flash_attn_ext_f16<
+            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+    }
+    launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
+}
+
+#define DECL_FATTN_WMMA_F16_CASE(D, cols_per_block, KQ_acc_t)                         \
+    template void ggml_cuda_flash_attn_ext_wmma_f16_case                              \
+    <D, cols_per_block, KQ_acc_t>(ggml_backend_cuda_context & ctx, ggml_tensor * dst) \
+
+extern DECL_FATTN_WMMA_F16_CASE( 64, 16, float);
+extern DECL_FATTN_WMMA_F16_CASE( 80, 16, float);
+extern DECL_FATTN_WMMA_F16_CASE( 96, 16, float);
+extern DECL_FATTN_WMMA_F16_CASE(112, 16, float);
+extern DECL_FATTN_WMMA_F16_CASE(128, 16, float);
+extern DECL_FATTN_WMMA_F16_CASE(256, 16, float);
+
+extern DECL_FATTN_WMMA_F16_CASE( 64, 32, float);
+extern DECL_FATTN_WMMA_F16_CASE( 80, 32, float);
+extern DECL_FATTN_WMMA_F16_CASE( 96, 32, float);
+extern DECL_FATTN_WMMA_F16_CASE(112, 32, float);
+extern DECL_FATTN_WMMA_F16_CASE(128, 32, float);
+// extern DECL_FATTN_WMMA_F16_CASE(256, 16, float);
+
+extern DECL_FATTN_WMMA_F16_CASE( 64,  8, half);
+extern DECL_FATTN_WMMA_F16_CASE( 96,  8, half);
+extern DECL_FATTN_WMMA_F16_CASE(128,  8, half);
+extern DECL_FATTN_WMMA_F16_CASE(256,  8, half);
+
+extern DECL_FATTN_WMMA_F16_CASE( 64, 16, half);
+extern DECL_FATTN_WMMA_F16_CASE( 80, 16, half);
+extern DECL_FATTN_WMMA_F16_CASE( 96, 16, half);
+extern DECL_FATTN_WMMA_F16_CASE(112, 16, half);
+extern DECL_FATTN_WMMA_F16_CASE(128, 16, half);
+extern DECL_FATTN_WMMA_F16_CASE(256, 16, half);
+
+extern DECL_FATTN_WMMA_F16_CASE( 64, 32, half);
+extern DECL_FATTN_WMMA_F16_CASE( 80, 32, half);
+extern DECL_FATTN_WMMA_F16_CASE( 96, 32, half);
+extern DECL_FATTN_WMMA_F16_CASE(112, 32, half);
+extern DECL_FATTN_WMMA_F16_CASE(128, 32, half);
+extern DECL_FATTN_WMMA_F16_CASE(256, 16, half);

llama/ggml-cuda/fattn.cu

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-tile-f16.cuh"
+#include "fattn-tile-f32.cuh"
+#include "fattn-vec-f16.cuh"
+#include "fattn-vec-f32.cuh"
+#include "fattn-wmma-f16.cuh"
+#include "fattn.cuh"
+
+#include <cstdint>
+
+static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    const int32_t precision = KQV->op_params[3];
+
+    if (precision != GGML_PREC_DEFAULT) {
+        if (Q->ne[1] <= 32 || Q->ne[0] > 128) {
+            constexpr int cols_per_block = 16;
+            switch (Q->ne[0]) {
+                case 64:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case< 64, cols_per_block, float>(ctx, dst);
+                    break;
+                case 80:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case< 80, cols_per_block, float>(ctx, dst);
+                    break;
+                case 96:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case< 96, cols_per_block, float>(ctx, dst);
+                    break;
+                case 112:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case<112, cols_per_block, float>(ctx, dst);
+                    break;
+                case 128:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, float>(ctx, dst);
+                    break;
+                case 256:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, float>(ctx, dst);
+                    break;
+                default:
+                    GGML_ABORT("fatal error");
+                    break;
+            }
+        } else {
+            constexpr int cols_per_block = 32;
+            switch (Q->ne[0]) {
+                case 64:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case< 64, cols_per_block, float>(ctx, dst);
+                    break;
+                case 80:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case< 80, cols_per_block, float>(ctx, dst);
+                    break;
+                case 96:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case< 96, cols_per_block, float>(ctx, dst);
+                    break;
+                case 112:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case<112, cols_per_block, float>(ctx, dst);
+                    break;
+                case 128:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, float>(ctx, dst);
+                    break;
+                // case 256:
+                //     ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, float>(ctx, dst);
+                //     break;
+                default:
+                    GGML_ABORT("fatal error");
+                    break;
+            }
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 8 && Q->ne[0] % WARP_SIZE == 0) {
+        constexpr int cols_per_block = 8;
+        switch (Q->ne[0]) {
+            case 64:
+                ggml_cuda_flash_attn_ext_wmma_f16_case< 64, cols_per_block, half>(ctx, dst);
+                break;
+            case 96:
+                ggml_cuda_flash_attn_ext_wmma_f16_case< 96, cols_per_block, half>(ctx, dst);
+                break;
+            case 128:
+                ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, half>(ctx, dst);
+                break;
+            case 256:
+                ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, half>(ctx, dst);
+                break;
+            default:
+                GGML_ABORT("fatal error");
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 32) {
+        constexpr int cols_per_block = 16;
+        switch (Q->ne[0]) {
+            case 64:
+                ggml_cuda_flash_attn_ext_wmma_f16_case< 64, cols_per_block, half>(ctx, dst);
+                break;
+            case 80:
+                ggml_cuda_flash_attn_ext_wmma_f16_case< 80, cols_per_block, half>(ctx, dst);
+                break;
+            case 96:
+                ggml_cuda_flash_attn_ext_wmma_f16_case< 96, cols_per_block, half>(ctx, dst);
+                break;
+            case 112:
+                ggml_cuda_flash_attn_ext_wmma_f16_case<112, cols_per_block, half>(ctx, dst);
+                break;
+            case 128:
+                ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, half>(ctx, dst);
+                break;
+            case 256:
+                ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, half>(ctx, dst);
+                break;
+            default:
+                GGML_ABORT("fatal error");
+                break;
+        }
+        return;
+    }
+
+    constexpr int cols_per_block = 32;
+    switch (Q->ne[0]) {
+        case 64:
+            ggml_cuda_flash_attn_ext_wmma_f16_case< 64, cols_per_block, half>(ctx, dst);
+            break;
+        case 80:
+            ggml_cuda_flash_attn_ext_wmma_f16_case< 80, cols_per_block, half>(ctx, dst);
+            break;
+        case 96:
+            ggml_cuda_flash_attn_ext_wmma_f16_case< 96, cols_per_block, half>(ctx, dst);
+            break;
+        case 112:
+            ggml_cuda_flash_attn_ext_wmma_f16_case<112, cols_per_block, half>(ctx, dst);
+            break;
+        case 128:
+            ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, half>(ctx, dst);
+            break;
+        case 256:
+            ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, half>(ctx, dst);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+#define FATTN_VEC_F16_CASE(D, type_K, type_V)                               \
+    if (Q->ne[0] == (D) && K->type == (type_K) && V->type == (type_V)) {    \
+        ggml_cuda_flash_attn_ext_vec_f16_case<D, type_K, type_V>(ctx, dst); \
+        return;                                                             \
+    }                                                                       \
+
+static void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * Q = dst->src[1];
+    ggml_tensor * K = dst->src[1];
+    ggml_tensor * V = dst->src[2];
+
+#ifdef GGML_CUDA_FA_ALL_QUANTS
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16 )
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_0)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_1)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_0)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_1)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q8_0)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_F16)
+
+    FATTN_VEC_F16_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
+#else
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
+
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
+#endif // GGML_CUDA_FA_ALL_QUANTS
+
+    on_no_fattn_vec_case(Q->ne[0]);
+}
+
+#define FATTN_VEC_F32_CASE(D, type_K, type_V)                               \
+    if (Q->ne[0] == (D) && K->type == (type_K) && V->type == (type_V)) {    \
+        ggml_cuda_flash_attn_ext_vec_f32_case<D, type_K, type_V>(ctx, dst); \
+        return;                                                             \
+    }                                                                       \
+
+static void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * Q = dst->src[1];
+    ggml_tensor * K = dst->src[1];
+    ggml_tensor * V = dst->src[2];
+
+#ifdef GGML_CUDA_FA_ALL_QUANTS
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_0)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_1)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_0)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_1)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q8_0)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_F16)
+
+    FATTN_VEC_F32_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
+#else
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
+
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
+#endif // GGML_CUDA_FA_ALL_QUANTS
+
+    on_no_fattn_vec_case(Q->ne[0]);
+}
+
+void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    ggml_cuda_set_device(ctx.device);
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const int32_t precision = KQV->op_params[3];
+
+    // On AMD the tile kernels perform poorly, use the vec kernel instead:
+    if (cc >= CC_OFFSET_AMD) {
+        if (precision == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
+            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        }
+        return;
+    }
+
+    if (!fast_fp16_available(cc)) {
+        if (Q->ne[1] <= 8) {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
+        }
+        return;
+    }
+
+    if (!fp16_mma_available(cc)) {
+        if (Q->ne[1] <= 8) {
+            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_tile_f16(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] == 1 && Q->ne[0] % (2*WARP_SIZE) == 0) {
+        if (precision == GGML_PREC_DEFAULT) {
+            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+            return;
+        } else if(Q->ne[0] <= 128) {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+            return;
+        }
+    }
+
+    ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
+}

llama/ggml-cuda/fattn.cuh

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/getrows.cu

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "getrows.cuh"
+#include "dequantize.cuh"
+
+template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static __global__ void k_get_rows(
+            const void * src0, const int32_t * src1, dst_t * dst,
+            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
+            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
+            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
+            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
+            size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
+
+    const int i00 = (blockIdx.x*blockDim.x + threadIdx.x)*2;
+    const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
+    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
+    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
+
+    if (i00 >= ne00) {
+        return;
+    }
+
+    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
+
+    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+    const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03;
+
+    const int ib = i00/qk; // block index
+    const int iqs = (i00%qk)/qr; // quant index
+    const int iybs = i00 - i00%qk; // dst block start index
+    const int y_offset = qr == 1 ? 1 : qk/2;
+
+    // dequantize
+    dfloat2 v;
+    dequantize_kernel(src0_row, ib, iqs, v);
+
+    dst_row[iybs + iqs + 0]        = v.x;
+    dst_row[iybs + iqs + y_offset] = v.y;
+}
+
+template<typename src0_t, typename dst_t>
+static __global__ void k_get_rows_float(
+            const src0_t * src0, const int32_t * src1, dst_t * dst,
+            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
+            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
+            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
+            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
+            size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
+
+    const int i00 = blockIdx.x*blockDim.x + threadIdx.x;
+    const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
+    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
+    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
+
+    if (i00 >= ne00) {
+        return;
+    }
+
+    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
+
+    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+    const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03);
+
+    dst_row[i00] = src0_row[i00];
+}
+
+template<int qk, int qr, dequantize_kernel_t dq>
+static void get_rows_cuda(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+                            const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
+    const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
+    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
+
+    // strides in elements
+    //const size_t s0 = nb0 / ggml_element_size(dst);
+    const size_t s1 = nb1 / ggml_element_size(dst);
+    const size_t s2 = nb2 / ggml_element_size(dst);
+    const size_t s3 = nb3 / ggml_element_size(dst);
+
+    const size_t s10 = nb10 / ggml_element_size(src1);
+    const size_t s11 = nb11 / ggml_element_size(src1);
+    const size_t s12 = nb12 / ggml_element_size(src1);
+    //const size_t s13 = nb13 / ggml_element_size(src1);
+
+    GGML_ASSERT(ne00 % 2 == 0);
+
+    k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
+            src0_dd, src1_dd, dst_dd,
+            ne00, /*ne01, ne02, ne03,*/
+            /*ne10, ne11,*/ ne12, /*ne13,*/
+            /* s0,*/ s1, s2, s3,
+            /* nb00,*/ nb01, nb02, nb03,
+            s10, s11, s12/*, s13*/);
+
+    GGML_UNUSED(dst);
+}
+
+template<typename src0_t>
+static void get_rows_cuda_float(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+                                const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
+    const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
+    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
+
+    // strides in elements
+    //const size_t s0 = nb0 / ggml_element_size(dst);
+    const size_t s1 = nb1 / ggml_element_size(dst);
+    const size_t s2 = nb2 / ggml_element_size(dst);
+    const size_t s3 = nb3 / ggml_element_size(dst);
+
+    const size_t s10 = nb10 / ggml_element_size(src1);
+    const size_t s11 = nb11 / ggml_element_size(src1);
+    const size_t s12 = nb12 / ggml_element_size(src1);
+    //const size_t s13 = nb13 / ggml_element_size(src1);
+
+    k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
+            src0_dd, src1_dd, dst_dd,
+            ne00, /*ne01, ne02, ne03,*/
+            /*ne10, ne11,*/ ne12, /*ne13,*/
+            /* s0,*/ s1, s2, s3,
+            /* nb00,*/ nb01, nb02, nb03,
+            s10, s11, s12/*, s13*/);
+
+    GGML_UNUSED(dst);
+}
+
+void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const float * src0_d = (const float *)src0->data;
+    const float * src1_d = (const float *)src1->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+
+    GGML_ASSERT(src1->type == GGML_TYPE_I32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
+    GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
+    GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type));
+
+    const int32_t * src1_i32 = (const int32_t *) src1_d;
+
+    switch (src0->type) {
+        case GGML_TYPE_F16:
+            get_rows_cuda_float(src0, src1, dst, (const half *)src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_F32:
+            get_rows_cuda_float(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q4_0:
+            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        default:
+            // TODO: k-quants
+            GGML_ABORT("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
+            break;
+    }
+}

llama/ggml-cuda/getrows.cuh

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "common.cuh"
+
+#define CUDA_GET_ROWS_BLOCK_SIZE 256
+
+void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/im2col.cu

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "im2col.cuh"
+
+template <typename T>
+static  __global__ void im2col_kernel(
+        const float * x, T * dst, int64_t batch_offset,
+        int64_t offset_delta, int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH, int64_t pelements, int64_t CHW,
+        int s0, int s1, int p0, int p1, int d0, int d1) {
+    const int64_t i = threadIdx.x + blockIdx.x * blockDim.x;
+    if (i >= pelements) {
+        return;
+    }
+
+    const int64_t  ksize = OW * (KH > 1 ? KW : 1);
+    const int64_t  kx = i / ksize;
+    const int64_t  kd = kx * ksize;
+    const int64_t  ky = (i - kd) / OW;
+    const int64_t  ix = i % OW;
+
+    const int64_t  oh = blockIdx.y;
+    const int64_t  batch = blockIdx.z / IC;
+    const int64_t  ic = blockIdx.z % IC;
+
+    const int64_t iiw = ix * s0 + kx * d0 - p0;
+    const int64_t iih = oh * s1 + ky * d1 - p1;
+
+    const int64_t offset_dst =
+        ((batch * OH + oh) * OW + ix) * CHW +
+        (ic * (KW * KH) + ky * KW + kx);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        dst[offset_dst] = 0.0f;
+    } else {
+        const int64_t offset_src = ic * offset_delta + batch * batch_offset;
+        dst[offset_dst] = x[offset_src + iih * IW + iiw];
+    }
+}
+
+template <typename T>
+static void im2col_cuda(const float * x, T* dst,
+    int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
+    int64_t batch, int64_t batch_offset, int64_t offset_delta,
+    int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
+    const int parallel_elements = OW * KW * KH;
+    const int num_blocks = (parallel_elements + CUDA_IM2COL_BLOCK_SIZE - 1) / CUDA_IM2COL_BLOCK_SIZE;
+    dim3 block_nums(num_blocks, OH, batch * IC);
+    im2col_kernel<<<block_nums, CUDA_IM2COL_BLOCK_SIZE, 0, stream>>>(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, (IC * KH * KW), s0, s1, p0, p1, d0, d1);
+}
+
+static void im2col_cuda_f16(const float * x, half * dst,
+    int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
+    int64_t batch, int64_t batch_offset, int64_t offset_delta,
+    int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
+
+    im2col_cuda<half>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0, p1, d0, d1, stream);
+}
+
+static void im2col_cuda_f32(const float * x, float * dst,
+    int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
+    int64_t batch, int64_t batch_offset, int64_t offset_delta,
+    int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
+
+    im2col_cuda<float>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0, p1, d0, d1, stream);
+}
+
+void ggml_cuda_op_im2col(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const float * src1_d = (const float *)src1->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
+
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
+
+    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
+
+    const int64_t IC = src1->ne[is_2D ? 2 : 1];
+    const int64_t IH = is_2D ? src1->ne[1] : 1;
+    const int64_t IW =         src1->ne[0];
+
+    const int64_t KH = is_2D ? src0->ne[1] : 1;
+    const int64_t KW =         src0->ne[0];
+
+    const int64_t OH = is_2D ? dst->ne[2] : 1;
+    const int64_t OW =         dst->ne[1];
+
+    const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
+    const int64_t batch = src1->ne[3];
+    const size_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32
+
+    if(dst->type == GGML_TYPE_F16) {
+        im2col_cuda_f16(src1_d, (half *) dst_d, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
+    } else {
+        im2col_cuda_f32(src1_d, (float *) dst_d, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
+    }
+}

llama/ggml-cuda/im2col.cuh

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "common.cuh"
+
+#define CUDA_IM2COL_BLOCK_SIZE 256
+
+void ggml_cuda_op_im2col(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/mma.cuh

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "common.cuh"
+
+struct mma_int_A_I16K4 {
+    static constexpr int I  = 16;
+    static constexpr int K  = 4;
+    static constexpr int ne = 2;
+
+    int x[ne] = {0};
+
+    static __device__ __forceinline__ int get_i(const int l) {
+        const int ret = (l%2) * (I/2) + threadIdx.x / K;
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  I);
+        return ret;
+    }
+
+    static __device__ __forceinline__ int get_k(const int /* l */) {
+        const int ret = threadIdx.x % K;
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  K);
+        return ret;
+    }
+
+    __device__ __forceinline__ void load(const int * __restrict__ xs0, const int & stride) {
+#if defined(INT8_MMA_AVAILABLE)
+        const int * xs = xs0 + (threadIdx.x%I)*stride;
+        asm("ldmatrix.sync.aligned.m8n8.x2.b16 {%0, %1}, [%2];"
+            : "+r"(x[0]), "+r"(x[1])
+            : "l"(xs));
+#else
+#pragma unroll
+        for (int l = 0; l < ne; ++l) {
+            x[l] = xs0[get_i(l)*stride + get_k(l)];
+        }
+#endif // defined(INT8_MMA_AVAILABLE)
+    }
+};
+
+struct mma_int_A_I16K8 {
+    static constexpr int I  = 16;
+    static constexpr int K  = 8;
+    static constexpr int ne = 4;
+
+    int x[ne] = {0};
+
+    static __device__ __forceinline__ int get_i(const int l) {
+        const int ret = (l%2) * (I/2) + threadIdx.x / (K/2);
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  I);
+        return ret;
+    }
+
+    static __device__ __forceinline__ int get_k(const int l) {
+        const int ret = (l/2) * (K/2) + threadIdx.x % (K/2);
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  K);
+        return ret;
+    }
+
+    __device__ __forceinline__ void load(const int * __restrict__ xs0, const int & stride) {
+#if defined(INT8_MMA_AVAILABLE)
+        const int * xs = xs0 + (threadIdx.x%I)*stride + (threadIdx.x/I)*(K/2);
+        asm("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];"
+            : "+r"(x[0]), "+r"(x[1]), "+r"(x[2]), "+r"(x[3])
+            : "l"(xs));
+#else
+#pragma unroll
+        for (int l = 0; l < ne; ++l) {
+            x[l] = xs0[get_i(l)*stride + get_k(l)];
+        }
+#endif // defined(INT8_MMA_AVAILABLE)
+    }
+
+    __device__ __forceinline__ void load_low(const int * __restrict__ xs0, const int & stride) {
+        ((mma_int_A_I16K4 *) x)[0].load(xs0, stride);
+    }
+};
+
+struct mma_int_B_J8K4 {
+    static constexpr int J  = 8;
+    static constexpr int K  = 4;
+    static constexpr int ne = 1;
+
+    int x[ne] = {0};
+
+    static __device__ __forceinline__ int get_j(const int /* l */) {
+        const int ret = threadIdx.x / K;
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  J);
+        return ret;
+    }
+
+    static __device__ __forceinline__ int get_k(const int /* l */) {
+        const int ret = threadIdx.x % K;
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  K);
+        return ret;
+    }
+
+    __device__ __forceinline__ void load(const int * __restrict__ xs0, const int & stride) {
+#if defined(INT8_MMA_AVAILABLE) && false // Loading as 4 byte values is faster
+        const int * xs = xs0 + (threadIdx.x%J)*stride;
+        asm("ldmatrix.sync.aligned.m8n8.x1.b16 {%0}, [%1];"
+            : "+r"(x[0])
+            : "l"(xs));
+#else
+#pragma unroll
+        for (int l = 0; l < ne; ++l) {
+            x[l] = xs0[get_j(l)*stride + get_k(l)];
+        }
+#endif // defined(INT8_MMA_AVAILABLE)
+    }
+};
+
+struct mma_int_B_J8K8 {
+    static constexpr int J  = 8;
+    static constexpr int K  = 8;
+    static constexpr int ne = 2;
+
+    int x[ne] = {0};
+
+    static __device__ __forceinline__ int get_j(const int /* l */) {
+        const int ret = threadIdx.x / (K/2);
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  J);
+        return ret;
+    }
+
+    static __device__ __forceinline__ int get_k(const int l) {
+        const int ret = l * (K/2) + threadIdx.x % (K/2);
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  K);
+        return ret;
+    }
+
+    __device__ __forceinline__ void load(const int * __restrict__ xs0, const int & stride) {
+#if defined(INT8_MMA_AVAILABLE) && false // Loading as 4 byte values is faster
+        const int * xs = xs0 + (threadIdx.x%J)*stride + ((threadIdx.x/J)*(K/2)) % K;
+        asm("ldmatrix.sync.aligned.m8n8.x2.b16 {%0, %1}, [%2];"
+            : "+r"(x[0]), "+r"(x[1])
+            : "l"(xs));
+#else
+#pragma unroll
+        for (int l = 0; l < ne; ++l) {
+            x[l] = xs0[get_j(l)*stride + get_k(l)];
+        }
+#endif // defined(INT8_MMA_AVAILABLE)
+    }
+};
+
+struct mma_int_C_I16J8 {
+    static constexpr int I  = 16;
+    static constexpr int J  = 8;
+    static constexpr int ne = 4;
+
+    int x[ne] = {0};
+
+    static __device__ __forceinline__ int get_i(const int l) {
+        const int ret = (l/2) * (I/2) + threadIdx.x / (J/2);
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  I);
+        return ret;
+    }
+
+    static __device__ __forceinline__ int get_j(const int l) {
+        const int ret = 2 * (threadIdx.x % (J/2)) + l%2;
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  J);
+        return ret;
+    }
+
+    __device__ __forceinline__ void mma_K4(const mma_int_A_I16K4 & mma_A, const mma_int_B_J8K4 & mma_B) {
+#ifdef INT8_MMA_AVAILABLE
+#if __CUDA_ARCH__ >= CC_AMPERE
+        asm("mma.sync.aligned.m16n8k16.row.col.s32.s8.s8.s32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};"
+            : "+r"(x[0]), "+r"(x[1]), "+r"(x[2]), "+r"(x[3])
+            : "r"(mma_A.x[0]), "r"(mma_A.x[1]), "r"(mma_B.x[0]));
+#else
+        // On Turing m16n8k16 mma is not available, use 2x m8n8k16 mma instead:
+        asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
+            : "+r"(x[0]), "+r"(x[1])
+            : "r"(mma_A.x[0]), "r"(mma_B.x[0]));
+        asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
+            : "+r"(x[2]), "+r"(x[3])
+            : "r"(mma_A.x[1]), "r"(mma_B.x[0]));
+#endif // __CUDA_ARCH__ >= CC_AMPERE
+#else
+        GGML_UNUSED(mma_A);
+        GGML_UNUSED(mma_B);
+        NO_DEVICE_CODE;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    __device__ __forceinline__ void mma_K8(const mma_int_A_I16K8 & mma_A, const mma_int_B_J8K8 & mma_B) {
+#ifdef INT8_MMA_AVAILABLE
+#if __CUDA_ARCH__ >= CC_AMPERE
+        asm("mma.sync.aligned.m16n8k32.row.col.s32.s8.s8.s32 {%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3};"
+            : "+r"(x[0]), "+r"(x[1]), "+r"(x[2]), "+r"(x[3])
+            : "r"(mma_A.x[0]), "r"(mma_A.x[1]), "r"(mma_A.x[2]), "r"(mma_A.x[3]), "r"(mma_B.x[0]), "r"(mma_B.x[1]));
+#else
+        // On Turing m16n8k32 mma is not available, use 4x m8n8k16 mma instead:
+        asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
+            : "+r"(x[0]), "+r"(x[1])
+            : "r"(mma_A.x[0]), "r"(mma_B.x[0]));
+        asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
+            : "+r"(x[2]), "+r"(x[3])
+            : "r"(mma_A.x[1]), "r"(mma_B.x[0]));
+        asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
+            : "+r"(x[0]), "+r"(x[1])
+            : "r"(mma_A.x[2]), "r"(mma_B.x[1]));
+        asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
+            : "+r"(x[2]), "+r"(x[3])
+            : "r"(mma_A.x[3]), "r"(mma_B.x[1]));
+#endif // __CUDA_ARCH__ >= CC_AMPERE
+#else
+        GGML_UNUSED(mma_A);
+        GGML_UNUSED(mma_B);
+        NO_DEVICE_CODE;
+#endif // INT8_MMA_AVAILABLE
+    }
+};

llama/ggml-cuda/mmq.cu

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "mmq.cuh"
+
+void ggml_cuda_op_mul_mat_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    const int64_t ne00 = src0->ne[0];
+
+    const int64_t nb01 = src0->nb[1];
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    GGML_ASSERT(ne10 % QK8_1 == 0);
+
+    const int64_t ne0 = dst->ne[0];
+
+    const int64_t row_diff = row_high - row_low;
+    const int64_t stride00 = nb01 / ggml_type_size(src0->type);
+
+    int id = ggml_cuda_get_device();
+    const int compute_capability = ggml_cuda_info().devices[id].cc;
+
+    // the main device has a larger memory buffer to hold the results from all GPUs
+    // nrows_dst == nrows of the matrix that the kernel writes into
+    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
+
+    const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst};
+
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0:
+            mul_mat_q_case<GGML_TYPE_Q4_0>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            mul_mat_q_case<GGML_TYPE_Q4_1>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            mul_mat_q_case<GGML_TYPE_Q5_0>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            mul_mat_q_case<GGML_TYPE_Q5_1>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            mul_mat_q_case<GGML_TYPE_Q8_0>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            mul_mat_q_case<GGML_TYPE_Q2_K>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            mul_mat_q_case<GGML_TYPE_Q3_K>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            mul_mat_q_case<GGML_TYPE_Q4_K>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            mul_mat_q_case<GGML_TYPE_Q5_K>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            mul_mat_q_case<GGML_TYPE_Q6_K>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ2_XXS:
+            mul_mat_q_case<GGML_TYPE_IQ2_XXS>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ2_XS:
+            mul_mat_q_case<GGML_TYPE_IQ2_XS>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ2_S:
+            mul_mat_q_case<GGML_TYPE_IQ2_S>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ3_XXS:
+            mul_mat_q_case<GGML_TYPE_IQ3_XXS>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ3_S:
+            mul_mat_q_case<GGML_TYPE_IQ3_S>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ1_S:
+            mul_mat_q_case<GGML_TYPE_IQ1_S>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ4_XS:
+            mul_mat_q_case<GGML_TYPE_IQ4_XS>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ4_NL:
+            mul_mat_q_case<GGML_TYPE_IQ4_NL>(ctx, args, stream);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddf_i);
+}
+
+bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
+#ifdef GGML_CUDA_FORCE_CUBLAS
+    return false;
+#endif // GGML_CUDA_FORCE_CUBLAS
+
+    bool mmq_supported;
+
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_S:
+        case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ4_XS:
+        case GGML_TYPE_IQ4_NL:
+            mmq_supported = true;
+            break;
+        default:
+            mmq_supported = false;
+            break;
+    }
+
+    if (!mmq_supported) {
+        return false;
+    }
+
+    if (int8_mma_available(cc)) {
+        return true;
+    }
+
+    if (cc < MIN_CC_DP4A) {
+        return false;
+    }
+
+#ifdef GGML_CUDA_FORCE_MMQ
+    return true;
+#endif //GGML_CUDA_FORCE_MMQ
+
+    if (cc < CC_OFFSET_AMD) {
+        return cc < CC_VOLTA || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+    }
+
+    return cc < CC_RDNA3 || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+}

llama/ggml-cuda/mmq.cuh

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include "common.cuh"
+#include "vecdotq.cuh"
+#include "mma.cuh"
+
+#include <climits>
+#include <cstdint>
+
+#define MMQ_DP4A_MAX_BATCH_SIZE 64 // Max. batch size to use for dp4a MMQ kernels when FP16 tensor cores are available.
+#define MMQ_ITER_K 256
+#define MMQ_NWARPS 8
+
+typedef void (*load_tiles_mmq_t)(const char * __restrict__ x, int * x_tile, const int & kbx0, const int & i_max, const int & stride);
+typedef void (*vec_dot_mmq_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00);
+typedef void (*mmq_write_back_t)(const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max);
+
+enum mmq_q8_1_ds_layout {
+    MMQ_Q8_1_DS_LAYOUT_D4,
+    MMQ_Q8_1_DS_LAYOUT_DS4,
+    MMQ_Q8_1_DS_LAYOUT_D2S6,
+};
+
+struct block_q8_1_mmq {
+    // The y float data is converted to a data layout that can simply be copied to shared memory as a contiguous block.
+    // The y float data is first grouped as blocks of 128 values.
+    // These blocks are then treated as individual data values and transposed.
+    //
+    // To avoid shared memory bank conflicts each block is padded with 16 bytes.
+    // This padding is also used to store block scales/partial sums.
+    // The scales multiplied with the quantized data are equal to the unquantized values.
+    // The partial sums are obtained by summing up a subgroup of the contained values (prior to quantization)
+    //     and are only needed for performance reasons.
+    //
+    // The exact data stored depends on the x data type.
+    union {
+        float d4[4];    // 1 32 bit scale per 32 values, stored as d0,d1,d2,d3
+        half2 ds4[4];   // 1 16 bit scale + 1 16 bit partial sum per 32 values, stored as d0,s0,d1,s1,d2,s2,d3,s3
+        half  d2s6[8];  // 1 16 bit scale per 64 values + 1 16 bit partial sum per 16 values for the first 96 values,
+                        //     stored as d0,d1,s1,s2,s3,s4,s5
+    };
+    int8_t qs[4*QK8_1]; // 128 values quantized to 8 bit each
+};
+static_assert(sizeof(block_q8_1_mmq) == 4*QK8_1 + 4*sizeof(half2), "Unexpected block_q8_1_mmq size");
+static_assert(sizeof(block_q8_1_mmq) == 4*sizeof(block_q8_1),      "Unexpected block_q8_1_mmq size");
+
+static mmq_q8_1_ds_layout mmq_get_q8_1_ds_layout(const ggml_type type_x) {
+    switch (type_x) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+            return MMQ_Q8_1_DS_LAYOUT_DS4;
+        case GGML_TYPE_Q5_0:
+            return MMQ_Q8_1_DS_LAYOUT_D4;
+        case GGML_TYPE_Q5_1:
+            return MMQ_Q8_1_DS_LAYOUT_DS4;
+        case GGML_TYPE_Q8_0:
+            return MMQ_Q8_1_DS_LAYOUT_D4;
+        case GGML_TYPE_Q2_K:
+            return MMQ_Q8_1_DS_LAYOUT_D2S6;
+        case GGML_TYPE_Q3_K:
+            return MMQ_Q8_1_DS_LAYOUT_D4;
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+            return MMQ_Q8_1_DS_LAYOUT_DS4;
+        case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_S:
+            return MMQ_Q8_1_DS_LAYOUT_D4;
+        case GGML_TYPE_IQ1_S:
+            return MMQ_Q8_1_DS_LAYOUT_DS4;
+        case GGML_TYPE_IQ4_XS:
+        case GGML_TYPE_IQ4_NL:
+            return MMQ_Q8_1_DS_LAYOUT_D4;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
+struct tile_x_sizes {
+    int qs;
+    int dm;
+    int sc;
+};
+
+static constexpr int get_mmq_x_max_host(const int cc) {
+    return int8_mma_available(cc) ? 128 :
+#ifdef GGML_CUDA_FORCE_MMQ
+        cc >= CC_VOLTA && cc < CC_OFFSET_AMD ? 128                     : 64;
+#else
+        cc >= CC_VOLTA && cc < CC_OFFSET_AMD ? MMQ_DP4A_MAX_BATCH_SIZE : 64;
+#endif // GGML_CUDA_FORCE_MMQ
+}
+
+static constexpr __device__ int get_mmq_x_max_device() {
+#ifdef INT8_MMA_AVAILABLE
+    return 128;
+#else // INT8_MMA_AVAILABLE
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    return 128;
+#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+
+#if __CUDA_ARCH__ >= CC_VOLTA
+#ifdef GGML_CUDA_FORCE_MMQ
+    return MMQ_DP4A_MAX_BATCH_SIZE;
+#else // GGML_CUDA_FORCE_MMQ
+    return 128;
+#endif // GGML_CUDA_FORCE_MMQ
+#else // __CUDA_ARCH__ >= CC_VOLTA
+
+    return 64;
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // INT8_MMA_AVAILABLE
+}
+
+static constexpr int get_mmq_y_host(const int cc) {
+    return cc >= CC_OFFSET_AMD ? (cc == CC_RDNA1 ? 64 : 128) : (cc >= CC_VOLTA ? 128 : 64);
+}
+
+static constexpr __device__ int get_mmq_y_device() {
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA1)
+    return 64;
+#else
+    return 128;
+#endif // defined RDNA1
+#else
+#if __CUDA_ARCH__ >= CC_VOLTA
+    return 128;
+#else
+    return 64;
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+}
+
+#define MMQ_DP4A_TXS_Q4_0    tile_x_sizes{mmq_y*WARP_SIZE   + mmq_y, mmq_y*WARP_SIZE/QI4_0   + mmq_y/QI4_0,     0}
+#define MMQ_DP4A_TXS_Q4_1    tile_x_sizes{mmq_y*WARP_SIZE   + mmq_y, mmq_y*WARP_SIZE/QI4_1   + mmq_y/QI4_1,     0}
+#define MMQ_DP4A_TXS_Q8_0    tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE*2/QI8_0 + mmq_y/(QI8_0/2), 0}
+#define MMQ_DP4A_TXS_Q8_0_16 tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE*4/QI8_0 + mmq_y/(QI8_0/4), 0}
+#define MMQ_DP4A_TXS_Q8_1    tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE*2/QI8_1 + mmq_y/(QI8_1/2), 0}
+#define MMQ_DP4A_TXS_Q2_K    tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE         + mmq_y,           0}
+#define MMQ_DP4A_TXS_Q3_K    tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y,                                     mmq_y*WARP_SIZE/8 + mmq_y/8}
+#define MMQ_DP4A_TXS_Q4_K    tile_x_sizes{mmq_y*WARP_SIZE   + mmq_y, mmq_y*WARP_SIZE/QI4_K,                     mmq_y*WARP_SIZE/8 + mmq_y/8}
+#define MMQ_DP4A_TXS_Q5_K    tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE/QI5_K   + mmq_y/QI5_K,     mmq_y*WARP_SIZE/8 + mmq_y/8}
+#define MMQ_DP4A_TXS_Q6_K    tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE/QI6_K   + mmq_y/QI6_K,     mmq_y*WARP_SIZE/8 + mmq_y/8}
+
+static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml_type type, int mmq_y) {
+    return type == GGML_TYPE_Q4_0 ? MMQ_DP4A_TXS_Q4_0 :
+        type == GGML_TYPE_Q4_1    ? MMQ_DP4A_TXS_Q4_1 :
+        type == GGML_TYPE_Q5_0    ? MMQ_DP4A_TXS_Q8_0 :
+        type == GGML_TYPE_Q5_1    ? MMQ_DP4A_TXS_Q8_1 :
+        type == GGML_TYPE_Q8_0    ? MMQ_DP4A_TXS_Q8_0 :
+        type == GGML_TYPE_Q2_K    ? MMQ_DP4A_TXS_Q2_K :
+        type == GGML_TYPE_Q3_K    ? MMQ_DP4A_TXS_Q3_K :
+        type == GGML_TYPE_Q4_K    ? MMQ_DP4A_TXS_Q4_K :
+        type == GGML_TYPE_Q5_K    ? MMQ_DP4A_TXS_Q5_K :
+        type == GGML_TYPE_Q6_K    ? MMQ_DP4A_TXS_Q6_K :
+        type == GGML_TYPE_IQ2_XXS ? MMQ_DP4A_TXS_Q8_0 :
+        type == GGML_TYPE_IQ2_XS  ? MMQ_DP4A_TXS_Q8_0_16 :
+        type == GGML_TYPE_IQ2_S   ? MMQ_DP4A_TXS_Q8_0_16 :
+        type == GGML_TYPE_IQ3_XXS ? MMQ_DP4A_TXS_Q8_0 :
+        type == GGML_TYPE_IQ3_S   ? MMQ_DP4A_TXS_Q8_0 :
+        type == GGML_TYPE_IQ1_S   ? MMQ_DP4A_TXS_Q8_0 :
+        type == GGML_TYPE_IQ4_XS  ? MMQ_DP4A_TXS_Q8_0 :
+        type == GGML_TYPE_IQ4_NL  ? MMQ_DP4A_TXS_Q8_0 :
+        tile_x_sizes{0, 0, 0};
+}
+
+#define MMQ_MMA_TILE_X_K_Q8_0 (2*WARP_SIZE + 2*WARP_SIZE/QI8_0                 + 4)
+#define MMQ_MMA_TILE_X_K_Q8_1 (2*WARP_SIZE + 2*WARP_SIZE/QI8_0                 + 4)
+#define MMQ_MMA_TILE_X_K_Q2_K (2*WARP_SIZE + WARP_SIZE                         + 4)
+#define MMQ_MMA_TILE_X_K_Q3_K (2*WARP_SIZE + WARP_SIZE/2                       + 4)
+#define MMQ_MMA_TILE_X_K_Q6_K (2*WARP_SIZE + WARP_SIZE/QI6_K     + WARP_SIZE/8 + 7)
+
+static_assert(MMQ_MMA_TILE_X_K_Q8_0 % 8 == 4, "Wrong padding.");
+static_assert(MMQ_MMA_TILE_X_K_Q8_1 % 8 == 4, "Wrong padding.");
+static_assert(MMQ_MMA_TILE_X_K_Q2_K % 8 == 4, "Wrong padding.");
+static_assert(MMQ_MMA_TILE_X_K_Q3_K % 8 == 4, "Wrong padding.");
+static_assert(MMQ_MMA_TILE_X_K_Q6_K % 8 == 4, "Wrong padding.");
+
+static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
+    return type == GGML_TYPE_Q4_0 ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_Q4_1    ? MMQ_MMA_TILE_X_K_Q8_1 :
+        type == GGML_TYPE_Q5_0    ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_Q5_1    ? MMQ_MMA_TILE_X_K_Q8_1 :
+        type == GGML_TYPE_Q8_0    ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_Q2_K    ? MMQ_MMA_TILE_X_K_Q2_K :
+        type == GGML_TYPE_Q3_K    ? MMQ_MMA_TILE_X_K_Q3_K :
+        type == GGML_TYPE_Q4_K    ? MMQ_MMA_TILE_X_K_Q8_1 :
+        type == GGML_TYPE_Q5_K    ? MMQ_MMA_TILE_X_K_Q8_1 :
+        type == GGML_TYPE_Q6_K    ? MMQ_MMA_TILE_X_K_Q6_K :
+        type == GGML_TYPE_IQ2_XXS ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_IQ2_XS  ? MMQ_MMA_TILE_X_K_Q3_K :
+        type == GGML_TYPE_IQ2_S   ? MMQ_MMA_TILE_X_K_Q3_K :
+        type == GGML_TYPE_IQ3_XXS ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_IQ3_S   ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_IQ1_S   ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_IQ4_XS  ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_IQ4_NL  ? MMQ_MMA_TILE_X_K_Q8_0 :
+        0;
+}
+
+#define MMQ_TILE_Y_K (WARP_SIZE + WARP_SIZE/QI8_1)
+
+static int mmq_get_granularity_host(const int mmq_x, const int cc) {
+    return int8_mma_available(cc) && mmq_x >= 48 ? 16 : 8;
+}
+
+#ifdef INT8_MMA_AVAILABLE
+static constexpr __device__ int mmq_get_granularity_device(const int mmq_x) {
+    return mmq_x >= 48 ? 16 : 8;
+}
+#else
+static constexpr __device__ int mmq_get_granularity_device(const int /* mmq_x */) {
+    return 8;
+}
+#endif // INT8_MMA_AVAILABLE
+
+// ------------------------------------------------------------
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_0(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + 2*WARP_SIZE);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kbx  = threadIdx.x / QI4_0;
+    const int kqsx = threadIdx.x % QI4_0;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbx;
+        const int qs0 = get_int_b2(bxi->qs, kqsx);
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI4_0) + kqsx + 0]     = __vsubss4((qs0 >> 0) & 0x0F0F0F0F, 0x08080808);
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI4_0) + kqsx + QI4_0] = __vsubss4((qs0 >> 4) & 0x0F0F0F0F, 0x08080808);
+#else
+        x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = qs0;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI4_0;
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_0) {
+        int i = i0 + threadIdx.y * QI4_0 + threadIdx.x / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbxd;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0       + kbxd] = bxi->d;
+#else
+        x_df[i*(WARP_SIZE/QI4_0) + i/QI4_0 + kbxd] = bxi->d;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q4_0_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + txs.qs;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_ds = (const half2 *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR4_0*VDR_Q4_0_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const int kyqs = QI8_1 * ((k01/2) / (QI8_1/2)) + (k01/2) % (QI8_1/2);
+
+                int u[2*VDR_Q4_0_Q8_1_MMQ];
+
+#pragma unroll
+                for (int l = 0; l < VDR_Q4_0_Q8_1_MMQ; ++l) {
+                    u[2*l+0] = y_qs[j*MMQ_TILE_Y_K + kyqs +  l];
+                    u[2*l+1] = y_qs[j*MMQ_TILE_Y_K + kyqs + (l + QI4_0)];
+                }
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMQ>
+                    (&x_qs[i*(WARP_SIZE + 1) + k0/QR4_0], u,
+                     x_df[i*(WARP_SIZE/QI4_0) + i/QI4_0 + k0/(QR4_0*QI4_0)], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_1(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + 2*WARP_SIZE);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kbx  = threadIdx.x / QI4_1;
+    const int kqsx = threadIdx.x % QI4_1;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbx;
+        const int qs0 = get_int_b4(bxi->qs, kqsx);
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI4_1) + kqsx + 0]     = (qs0 >> 0) & 0x0F0F0F0F;
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI4_1) + kqsx + QI4_1] = (qs0 >> 4) & 0x0F0F0F0F;
+#else
+        x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = qs0;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI4_1;
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_1) {
+        int i = i0 + threadIdx.y * QI4_1 + threadIdx.x / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbxd;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_dm[i*MMQ_MMA_TILE_X_K_Q8_1       + kbxd] = bxi->dm;
+#else
+        x_dm[i*(WARP_SIZE/QI4_1) + i/QI4_1 + kbxd] = bxi->dm;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q4_1_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const half2 * x_dm = (const half2 *) x_qs + txs.qs;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_ds = (const half2 *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR4_1*VDR_Q4_1_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const int kyqs = QI8_1 * ((k01/2) / (QI8_1/2)) + (k01/2) % (QI8_1/2);
+
+                int u[2*VDR_Q4_1_Q8_1_MMQ];
+
+#pragma unroll
+                for (int l = 0; l < VDR_Q4_1_Q8_1_MMQ; ++l) {
+                    u[2*l+0] = y_qs[j*MMQ_TILE_Y_K + kyqs +  l];
+                    u[2*l+1] = y_qs[j*MMQ_TILE_Y_K + kyqs + (l + QI4_1)];
+                }
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMQ>
+                    (&x_qs[i*(WARP_SIZE + 1) + k0/QR4_1], u,
+                     x_dm[i*(WARP_SIZE/QI4_1) + i/QI4_1 + k0/(QR4_1*QI4_1)], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_0(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_0, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kbx  = threadIdx.x / QI5_0;
+    const int kqsx = threadIdx.x % QI5_0;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbx;
+
+        const int ql = get_int_b2(bxi->qs, kqsx);
+        const int qh = get_int_b2(bxi->qh, 0) >> (4 * (threadIdx.x % QI5_0));
+
+        int qs0 = (ql >>  0)   & 0x0F0F0F0F;
+        qs0    |= (qh <<  4)   & 0x00000010;  // 0 ->  4
+        qs0    |= (qh << 11)   & 0x00001000;  // 1 -> 12
+        qs0    |= (qh << 18)   & 0x00100000;  // 2 -> 20
+        qs0    |= (qh << 25)   & 0x10000000;  // 3 -> 28
+        qs0     = __vsubss4(qs0, 0x10101010); // subtract 16
+
+        int qs1 = (ql >>  4)   & 0x0F0F0F0F;
+        qs1    |= (qh >> 12)   & 0x00000010;  // 16 ->  4
+        qs1    |= (qh >>  5)   & 0x00001000;  // 17 -> 12
+        qs1    |= (qh <<  2)   & 0x00100000;  // 18 -> 20
+        qs1    |= (qh <<  9)   & 0x10000000;  // 19 -> 28
+        qs1     = __vsubss4(qs1, 0x10101010); // subtract 16
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI5_0) + kqsx + 0]     = qs0;
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1;
+#else
+        x_qs[i*(2*WARP_SIZE + 1)     + kbx*(2*QI5_0) + kqsx + 0]     = qs0;
+        x_qs[i*(2*WARP_SIZE + 1)     + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI5_0;
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_0) {
+        int i = i0 + threadIdx.y * QI5_0 + threadIdx.x / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbxd;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0       + kbxd] = bxi->d;
+#else
+        x_df[i*(WARP_SIZE/QI5_0) + i/QI5_0 + kbxd] = bxi->d;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_1(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + 2*WARP_SIZE);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kbx  = threadIdx.x / QI5_1;
+    const int kqsx = threadIdx.x % QI5_1;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbx;
+
+        const int ql = get_int_b4(bxi->qs, kqsx);
+        const int qh = get_int_b4(bxi->qh, 0) >> (4 * (threadIdx.x % QI5_1));
+
+        int qs0 = (ql >>  0) & 0x0F0F0F0F;
+        qs0    |= (qh <<  4) & 0x00000010; // 0 ->  4
+        qs0    |= (qh << 11) & 0x00001000; // 1 -> 12
+        qs0    |= (qh << 18) & 0x00100000; // 2 -> 20
+        qs0    |= (qh << 25) & 0x10000000; // 3 -> 28
+
+        int qs1 = (ql >>  4) & 0x0F0F0F0F;
+        qs1    |= (qh >> 12) & 0x00000010; // 16 ->  4
+        qs1    |= (qh >>  5) & 0x00001000; // 17 -> 12
+        qs1    |= (qh <<  2) & 0x00100000; // 18 -> 20
+        qs1    |= (qh <<  9) & 0x10000000; // 19 -> 28
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI5_1) + kqsx + 0]     = qs0;
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1;
+#else
+        x_qs[i*(2*WARP_SIZE + 1)     + kbx*(2*QI5_1) + kqsx + 0]     = qs0;
+        x_qs[i*(2*WARP_SIZE + 1)     + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI5_1;
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_1) {
+        int i = i0 + threadIdx.y * QI5_1 + threadIdx.x / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbxd;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_dm[i*MMQ_MMA_TILE_X_K_Q8_1       + kbxd] = bxi->dm;
+#else
+        x_dm[i*(WARP_SIZE/QI5_1) + i/QI5_1 + kbxd] = bxi->dm;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q8_0(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_tile + 2*WARP_SIZE);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kbx  = threadIdx.x / QI8_0;
+    const int kqsx = threadIdx.x % QI8_0;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbx;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 0         + threadIdx.x] = get_int_b2(bxi[0].qs,               kqsx);
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + WARP_SIZE + threadIdx.x] = get_int_b2(bxi[WARP_SIZE/QI8_0].qs, kqsx);
+#else
+        x_qs[i*(2*WARP_SIZE + 1)     + 0         + threadIdx.x] = get_int_b2(bxi[0].qs,               kqsx);
+        x_qs[i*(2*WARP_SIZE + 1)     + WARP_SIZE + threadIdx.x] = get_int_b2(bxi[WARP_SIZE/QI8_0].qs, kqsx);
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    const int blocks_per_tile_x_row = 2*WARP_SIZE / QI8_0;
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI8_0/2) {
+        int i = i0 + threadIdx.y * (QI8_0/2) + threadIdx.x / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbxd;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0             + kbxd] = bxi->d;
+#else
+        x_df[i*(2*WARP_SIZE/QI8_0) + i/(QI8_0/2) + kbxd] = bxi->d;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + txs.qs;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += VDR_Q8_0_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q8_0_q8_1_impl<float, VDR_Q8_0_Q8_1_MMQ>
+                    (&x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k0 % WARP_SIZE],
+                     x_df[i*(2*WARP_SIZE/QI8_0) + i/(QI8_0/2) + k0/QI8_0], y_df[j*MMQ_TILE_Y_K + (k0/QI8_1) % (WARP_SIZE/QI8_1)]);
+            }
+        }
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps, mmq_q8_1_ds_layout ds_layout>
+static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    typedef mma_int_A_I16K8 mma_A;
+    typedef mma_int_B_J8K8  mma_B;
+    typedef mma_int_C_I16J8 mma_C;
+
+    constexpr int granularity = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = 2 * granularity;
+    constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
+
+    y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
+
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + 2*WARP_SIZE;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+    const half2 * y_ds = (const half2 *) y;
+
+    mma_A A[ntx][WARP_SIZE/QI8_0];
+    float dA[ntx][mma_C::ne/2][WARP_SIZE/QI8_0];
+
+    const int i0 = (threadIdx.y/ntx)*rows_per_warp;
+
+#pragma unroll
+    for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_0) {
+            const int k0 = k00 + k01;
+
+            A[n][k01/QI8_0].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q8_0 + k0, MMQ_MMA_TILE_X_K_Q8_0);
+        }
+
+#pragma unroll
+        for (int l = 0; l < mma_C::ne/2; ++l) {
+            const int i = i0 + n*mma_A::I + mma_C::get_i(2*l);
+
+#pragma unroll
+            for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_0) {
+                const int k0 = k00 + k01;
+
+                dA[n][l][k01/QI8_0] = x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + k0/QI8_0];
+            }
+        }
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_0) {
+            mma_B  B;
+            float dB[mma_C::ne/2];
+
+            B.load(y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+
+#pragma unroll
+            for (int l = 0; l < mma_C::ne/2; ++l) {
+                const int j = j0 + mma_C::get_j(l);
+
+                if (ds_layout == MMQ_Q8_1_DS_LAYOUT_D4) {
+                    dB[l] =             y_df[j*MMQ_TILE_Y_K + k01/QI8_1];
+                } else {
+                    dB[l] = __low2float(y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]);
+                }
+            }
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+                mma_C C;
+                C.mma_K8(A[n][k01/QI8_0], B);
+
+#pragma unroll
+                for (int l = 0; l < mma_C::ne; ++l) {
+                    sum[(j0/mma_C::J + n)*mma_C::ne + l] += C.x[l]*dA[n][l/2][k01/QI8_0]*dB[l%2];
+                }
+            }
+        }
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q8_1_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const half2 * x_dm = (const half2 *) x_qs + txs.qs;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_ds = (const half2 *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += VDR_Q8_0_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q8_1_q8_1_impl<QR5_1*VDR_Q5_1_Q8_1_MMQ>
+                    (&x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01],
+                    x_dm[i*(WARP_SIZE/QI5_1) + i/QI5_1 + k0/QI8_1], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    typedef mma_int_A_I16K8 mma_A;
+    typedef mma_int_B_J8K8  mma_B;
+    typedef mma_int_C_I16J8 mma_C;
+
+    constexpr int granularity = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = 2 * granularity;
+    constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
+
+    y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
+
+    const int   * x_qs = (const int   *) x;
+    const half2 * x_dm = (const half2 *) x_qs + 2*WARP_SIZE;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_dm = (const half2 *) y;
+
+    mma_A    A[ntx][WARP_SIZE/QI8_1];
+    float2 dmA[ntx][mma_C::ne/2][WARP_SIZE/QI8_1];
+
+    const int i0 = (threadIdx.y/ntx)*rows_per_warp;
+
+#pragma unroll
+    for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1) {
+            const int k0 = k00 + k01;
+
+            A[n][k01/QI8_1].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q8_1 + k0, MMQ_MMA_TILE_X_K_Q8_1);
+        }
+
+#pragma unroll
+        for (int l = 0; l < mma_C::ne/2; ++l) {
+            const int i = i0 + n*mma_A::I + mma_C::get_i(2*l);
+
+#pragma unroll
+            for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1) {
+                const int k0 = k00 + k01;
+
+                dmA[n][l][k01/QI8_1] = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + k0/QI8_1]);
+            }
+        }
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1) {
+            mma_B    B;
+            float2 dsB[mma_C::ne/2];
+
+            B.load(y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+
+#pragma unroll
+            for (int l = 0; l < mma_C::ne/2; ++l) {
+                const int j = j0 + mma_C::get_j(l);
+
+                dsB[l] = __half22float2(y_dm[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+                mma_C C;
+                C.mma_K8(A[n][k01/QI8_1], B);
+
+#pragma unroll
+                for (int l = 0; l < mma_C::ne; ++l) {
+                    sum[(j0/mma_C::J + n)*mma_C::ne + l] += dmA[n][l/2][k01/QI8_1].x*dsB[l%2].x*C.x[l];
+                    sum[(j0/mma_C::J + n)*mma_C::ne + l] += dmA[n][l/2][k01/QI8_1].y*dsB[l%2].y;
+                }
+            }
+        }
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16;
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + txs.qs;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_0) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q8_0_16_q8_1_impl<QI8_0>(
+                    &x_qs[i*(2*WARP_SIZE + 1) + k0],
+                    &y_qs[j*MMQ_TILE_Y_K + k01],
+                    &x_df[i*(2*WARP_SIZE*2/QI8_0) + i/(QI8_0/4) + k0/(QI8_0/2)],
+                    y_df[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+#ifdef INT8_MMA_AVAILABLE
+
+    typedef mma_int_A_I16K4 mma_A;
+    typedef mma_int_A_I16K8 mma_A_K8;
+    typedef mma_int_B_J8K4  mma_B;
+    typedef mma_int_C_I16J8 mma_C;
+
+    constexpr int granularity = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = 2 * granularity;
+    constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
+
+    y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
+
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + WARP_SIZE*2;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+
+    const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I);
+
+    mma_A   A[ntx][8];
+    float  dA[ntx][mma_C::ne/2][8];
+
+#pragma unroll
+    for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += 8) {
+            const int k0 = k00 + k01;
+
+            ((mma_A_K8 *) A[n])[k01/8].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K);
+        }
+
+#pragma unroll
+        for (int l = 0; l < mma_C::ne/2; ++l) {
+            const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
+
+#pragma unroll
+            for (int k01 = 0; k01 < WARP_SIZE; k01 += 4) {
+                const int k0 = k00 + k01;
+
+                dA[n][l][k01/4] = x_df[i*MMQ_MMA_TILE_X_K_Q3_K + k0/4];
+            }
+        }
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += QR3_K*VDR_Q3_K_Q8_1_MMQ) {
+            mma_B B[2];
+            float dB[mma_C::ne/2];
+
+            B[0].load(y_qs + j0*MMQ_TILE_Y_K + (k01 + 0),        MMQ_TILE_Y_K);
+            B[1].load(y_qs + j0*MMQ_TILE_Y_K + (k01 + mma_B::K), MMQ_TILE_Y_K);
+
+#pragma unroll
+            for (int l = 0; l < mma_C::ne/2; ++l) {
+                const int j = j0 + mma_C::get_j(l);
+
+                dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1];
+            }
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+                mma_C C[2];
+                C[0].mma_K4(A[n][k01/4 + 0], B[0]);
+                C[1].mma_K4(A[n][k01/4 + 1], B[1]);
+
+#pragma unroll
+                for (int l = 0; l < mma_C::ne; ++l) {
+                    sum[(j0/mma_C::J + n)*mma_C::ne + l] += dB[l%2]*(C[0].x[l]*dA[n][l/2][k01/4 + 0] + C[1].x[l]*dA[n][l/2][k01/4 + 1]);
+                }
+            }
+        }
+    }
+#else
+    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
+    NO_DEVICE_CODE;
+#endif // INT8_MMA_AVAILABLE
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q2_K(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + 2*WARP_SIZE);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % QI2_K;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/QI2_K) {
+        int i = i0 + threadIdx.y*(WARP_SIZE/QI2_K) + threadIdx.x/QI2_K;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q2_K * bxi = (const block_q2_K *) x + kbx0 + i*stride;
+
+        const int x_ql_0 = get_int_b2(bxi->qs, kqsx);
+
+#pragma unroll
+        for (int l = 0; l < QR2_K; ++l) {
+            const int k = (kqsx/8)*32 + l*8 + kqsx % 8;
+
+            const int x_qs_k = (x_ql_0 >> (2*l)) & 0x03030303;
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q2_K + k] = x_qs_k;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + k] = x_qs_k;
+#endif // INT8_MMA_AVAILABLE
+        }
+
+        const int sc_m = bxi->scales[kqsx];
+#ifdef FAST_FP16_AVAILABLE
+        const half2 x_dm_ik = __hmul2(bxi->dm, make_half2(sc_m & 0x0F, sc_m >> 4));
+#else
+        const float2 bxi_dmf = __half22float2(bxi->dm);
+        const half2 x_dm_ik = make_half2(bxi_dmf.x*(sc_m & 0x0F), bxi_dmf.y*(sc_m >> 4));
+#endif // FAST_FP16_AVAILABLE
+
+#ifdef INT8_MMA_AVAILABLE
+        x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + kqsx] = x_dm_ik;
+#else
+        x_dm[i*(WARP_SIZE + 1)       + kqsx] = x_dm_ik;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const half2 * x_dm = (const half2 *) x_qs + txs.qs;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_ds = (const half2 *) y;
+
+    float2 y_df[mmq_x/nwarps];
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+        y_df[j0/nwarps] = __half22float2(y_ds[j*MMQ_TILE_Y_K]);
+    }
+
+#pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                if (k01 < WARP_SIZE/2) {
+                    constexpr int ns = 2;
+                    sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq<ns>(
+                        &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01],
+                        &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y,
+                        &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]);
+                } else {
+                    constexpr int ns = 1;
+                    sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq<ns>(
+                        &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01],
+                        &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y,
+                        &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]);
+                }
+            }
+        }
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+#ifdef INT8_MMA_AVAILABLE
+
+    typedef mma_int_A_I16K4 mma_A;
+    typedef mma_int_A_I16K8 mma_A_K8;
+    typedef mma_int_B_J8K4  mma_B;
+    typedef mma_int_C_I16J8 mma_C;
+
+    constexpr int granularity = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = 2 * granularity;
+    constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
+
+    y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
+
+    const int   * x_qs = (const int   *) x;
+    const half2 * x_dm = (const half2 *) x_qs + WARP_SIZE*2;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_ds = (const half2 *) y;
+
+    const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I);
+
+    mma_A   A[ntx][8];
+    float  dA[ntx][mma_C::ne/2][8];
+    float  mA[ntx][mma_C::ne/2][8];
+
+#pragma unroll
+    for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1) {
+            const int k0 = k00 + k01;
+
+            ((mma_A_K8 *) A[n])[k01/QI8_1].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K);
+        }
+    }
+
+#pragma unroll
+    for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+        for (int l = 0; l < mma_C::ne/2; ++l) {
+            const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
+
+#pragma unroll
+            for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1/2) {
+                const int k0 = k00 + k01;
+
+                const float2 dm = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + k0/(QI8_1/2)]);
+
+                dA[n][l][k01/(QI8_1/2)] = dm.x;
+                mA[n][l][k01/(QI8_1/2)] = dm.y;
+            }
+        }
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
+        float2 dB[mma_C::ne/2];
+
+#pragma unroll
+        for (int l = 0; l < mma_C::ne/2; ++l) {
+            const int j = j0 + mma_C::get_j(l);
+
+            dB[l] = __half22float2(y_ds[j*MMQ_TILE_Y_K]);
+        }
+
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1) {
+            mma_B B[2];
+
+            B[0].load(y_qs + j0*MMQ_TILE_Y_K + (k01 + 0),        MMQ_TILE_Y_K);
+            B[1].load(y_qs + j0*MMQ_TILE_Y_K + (k01 + mma_B::K), MMQ_TILE_Y_K);
+
+            mma_C Cm[2];
+            if (k01 >= WARP_SIZE * 3/4) {
+                mma_A A1;
+                A1.x[0] = 0x01010101;
+                A1.x[1] = 0x01010101;
+                Cm[0].mma_K4(A1, B[0]);
+                Cm[1].mma_K4(A1, B[1]);
+            }
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+                mma_C Cd[2];
+
+                Cd[0].mma_K4(A[n][k01/4 + 0], B[0]);
+                Cd[1].mma_K4(A[n][k01/4 + 1], B[1]);
+
+#pragma unroll
+                for (int l = 0; l < mma_C::ne; ++l) {
+                    float tmp = Cd[0].x[l]*dA[n][l/2][k01/4 + 0] + Cd[1].x[l]*dA[n][l/2][k01/4 + 1];
+                    if (k01 >= WARP_SIZE * 3/4) {
+                        tmp -= Cm[0].x[l]*mA[n][l/2][k01/4 + 0] + Cm[1].x[l]*mA[n][l/2][k01/4 + 1];
+                    }
+                    sum[(j0/mma_C::J + n)*mma_C::ne + l] += tmp*(k01 < WARP_SIZE/2 ? dB[l%2].x : dB[l%2].y);
+                }
+            }
+        }
+
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE * 3/4; k01 += QI8_1) {
+            float2 sB[mma_C::ne/2];
+
+#pragma unroll
+            for (int l = 0; l < mma_C::ne/2; ++l) {
+                const int j = j0 + mma_C::get_j(l);
+
+                sB[l] = __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]);
+            }
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+                for (int l = 0; l < mma_C::ne; ++l) {
+                    sum[(j0/mma_C::J + n)*mma_C::ne + l] -= mA[n][l/2][k01/4 + 0]*sB[l%2].x;
+                    sum[(j0/mma_C::J + n)*mma_C::ne + l] -= mA[n][l/2][k01/4 + 1]*sB[l%2].y;
+                }
+            }
+        }
+    }
+#else
+    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
+    NO_DEVICE_CODE;
+#endif // INT8_MMA_AVAILABLE
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q3_K(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+    int   * x_sc = (int   *) (x_df + txs.dm);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % QI3_K;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/QI3_K) {
+        int i = i0 + threadIdx.y * (WARP_SIZE/QI3_K) + threadIdx.x / QI3_K;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride;
+
+        const int x_ql_0 = get_int_b2(bxi->qs,    kqsx);
+        const int x_qh_0 = get_int_b2(bxi->hmask, kqsx % (QI3_K/2)) >> (4 * (kqsx / (QI3_K/2)));
+
+#pragma unroll
+        for (int l = 0; l < QR3_K; ++l) {
+            const int k = (kqsx/8)*32 + l*8 + kqsx % 8;
+
+            const int x_ql_k =  (x_ql_0 >> (2*l))       & 0x03030303;
+            const int x_qh_k = ((x_qh_0 >>    l)  << 2) & 0x04040404;
+
+            const int x_qs_k = __vsubss4(x_ql_k | x_qh_k, 0x04040404);
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + k] = x_qs_k;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + k] = x_qs_k;
+#endif // INT8_MMA_AVAILABLE
+        }
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps*8) {
+        int i = i0 + threadIdx.y*8 + threadIdx.x/(WARP_SIZE/8);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride;
+
+        const int ksc = threadIdx.x % (WARP_SIZE/8);
+
+        const int ksc_low = ksc % (QI3_K/8);
+        const int shift_low = 4 * (ksc / (QI3_K/8));
+        const int sc_low = (get_int_b2(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F;
+
+        const int ksc_high = QI3_K/8;
+        const int shift_high = 2 * ksc;
+        const int sc_high = ((get_int_b2(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030;
+
+        const int sc = __vsubss4(sc_low | sc_high, 0x20202020);
+
+#ifdef INT8_MMA_AVAILABLE
+        const int8_t * sc8 = (const int8_t *) &sc;
+        const float d = bxi->d;
+
+#pragma unroll
+        for (int l = 0; l < sizeof(int); ++l) {
+            x_df[i*MMQ_MMA_TILE_X_K_Q3_K + sizeof(int)*(threadIdx.x % (WARP_SIZE/8)) + l] = d*sc8[l];
+        }
+#else
+        x_sc[i*(WARP_SIZE/8) + i/8 + threadIdx.x % (WARP_SIZE/8)] = sc;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+#ifndef INT8_MMA_AVAILABLE
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps*WARP_SIZE) {
+        int i = (i0 + threadIdx.y*WARP_SIZE + threadIdx.x) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride;
+
+        x_df[i] = bxi->d;
+    }
+#endif // INT8_MMA_AVAILABLE
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q3_K_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + txs.qs;
+    const int   * x_sc = (const int   *) x_df + txs.dm;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR3_K*VDR_Q3_K_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const int8_t * scales = ((const int8_t *) (x_sc + i*(WARP_SIZE/8) + i/8)) + k0/4;
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q3_K_q8_1_impl_mmq(
+                    &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], scales,
+                    x_df[i], y_df[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+static __device__ __forceinline__ int unpack_scales_q45_K(const int * scales, const int ksc) {
+    // scale arrangement after the following two lines:
+    //   - ksc == 0: sc0, sc1, sc2, sc3
+    //   - ksc == 1: sc4, sc5, sc6, sc7
+    //   - ksc == 2:  m0,  m1,  m2,  m3
+    //   - ksc == 3:  m4,  m5,  m6,  m7
+    return ((scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F) | // lower 4 bits
+           ((scales[ksc/2]              >> (2 * (ksc % 2)))       & 0x30303030);  // upper 2 bits
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_K(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + 2*WARP_SIZE);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + txs.qs);
+    int   * x_sc = (int   *) (x_dm + txs.dm);
+#endif // INT8_MMA_AVAILABLE
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride;
+        const int qs0 = get_int_b4(bxi->qs, threadIdx.x);
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 16*(threadIdx.x/8) + threadIdx.x % 8 + 0] = (qs0 >> 0) & 0x0F0F0F0F;
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 16*(threadIdx.x/8) + threadIdx.x % 8 + 8] = (qs0 >> 4) & 0x0F0F0F0F;
+#else
+        x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = qs0;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+#ifdef INT8_MMA_AVAILABLE
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps*16) {
+        int i = (i0 + threadIdx.y*16 + threadIdx.x/(WARP_SIZE/16)) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride;
+
+        const int * scales = (const int *) bxi->scales;
+        const int ksc = threadIdx.x % (WARP_SIZE/16);
+
+        const int sc32 = unpack_scales_q45_K(scales, ksc + 0);
+        const int  m32 = unpack_scales_q45_K(scales, ksc + 2);
+
+        const uint8_t * sc8 = (const uint8_t *) &sc32;
+        const uint8_t *  m8 = (const uint8_t *)  &m32;
+
+        const half2 dm = bxi->dm * make_half2(1.0f, -1.0f);
+
+#pragma unroll
+        for (int l = 0; l < sizeof(int); ++l) {
+            x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]);
+        }
+    }
+
+#else
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps*QI4_K) {
+        int i = (i0 + threadIdx.y*QI4_K + threadIdx.x) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride;
+
+        x_dm[i] = bxi->dm;
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
+        int i = (i0 + threadIdx.y * 8 + threadIdx.x / (WARP_SIZE/8)) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride + (threadIdx.x % (WARP_SIZE/8)) / (QI4_K/8);
+
+        const int * scales = (const int *) bxi->scales;
+
+        const int ksc = threadIdx.x % (WARP_SIZE/8);
+        const int scales8 = unpack_scales_q45_K(scales, ksc);
+
+        x_sc[i*(WARP_SIZE/8) + i/8 + ksc] = scales8;
+    }
+#endif // INT8_MMA_AVAILABLE
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q4_K_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const half2 * x_dm = (const half2 *) x_qs + txs.qs;
+    const int   * x_sc = (const int   *) x_dm + txs.dm;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_ds = (const half2 *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR4_K*VDR_Q4_K_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const uint8_t * sc = (const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k0/32] + 2*(k01/16);
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q4_K_q8_1_impl_mmq(
+                    &x_qs[i*(WARP_SIZE + 1) + k0/2], &y_qs[j*MMQ_TILE_Y_K + k01], sc, sc+8,
+                    x_dm[i], &y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_K(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + txs.qs);
+    int   * x_sc = (int   *) (x_dm + txs.dm);
+#endif // INT8_MMA_AVAILABLE
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
+        const int ky = QR5_K*threadIdx.x;
+
+        const int ql = get_int_b4(bxi->qs, threadIdx.x);
+        const int ql0 = (ql >> 0) & 0x0F0F0F0F;
+        const int ql1 = (ql >> 4) & 0x0F0F0F0F;
+
+        const int qh = get_int_b4(bxi->qh, threadIdx.x % (QI5_K/4));
+        const int qh0 = ((qh >> (2 * (threadIdx.x / (QI5_K/4)) + 0)) << 4) & 0x10101010;
+        const int qh1 = ((qh >> (2 * (threadIdx.x / (QI5_K/4)) + 1)) << 4) & 0x10101010;
+
+        const int kq0 = ky - ky % (QI5_K/2) + threadIdx.x % (QI5_K/4) + 0;
+        const int kq1 = ky - ky % (QI5_K/2) + threadIdx.x % (QI5_K/4) + QI5_K/4;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kq0] = ql0 | qh0;
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kq1] = ql1 | qh1;
+#else
+        x_qs[i*(2*WARP_SIZE + 1)     + kq0] = ql0 | qh0;
+        x_qs[i*(2*WARP_SIZE + 1)     + kq1] = ql1 | qh1;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+#ifdef INT8_MMA_AVAILABLE
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps*16) {
+        int i = (i0 + threadIdx.y*16 + threadIdx.x/(WARP_SIZE/16)) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
+
+        const int * scales = (const int *) bxi->scales;
+        const int ksc = threadIdx.x % (WARP_SIZE/16);
+
+        const int sc32 = unpack_scales_q45_K(scales, ksc + 0);
+        const int  m32 = unpack_scales_q45_K(scales, ksc + 2);
+
+        const uint8_t * sc8 = (const uint8_t *) &sc32;
+        const uint8_t *  m8 = (const uint8_t *)  &m32;
+
+        const half2 dm = bxi->dm * make_half2(1.0f, -1.0f);
+
+#pragma unroll
+        for (int l = 0; l < sizeof(int); ++l) {
+            x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]);
+        }
+    }
+
+#else
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps*QI5_K) {
+        int i = (i0 + threadIdx.y*QI5_K + threadIdx.x) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
+
+        x_dm[i] = bxi->dm;
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps*8) {
+        int i = (i0 + threadIdx.y*8 + threadIdx.x/(WARP_SIZE/8)) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
+
+        const int * scales = (const int *) bxi->scales;
+
+        const int ksc = threadIdx.x % (WARP_SIZE/8);
+        const int scales8 = unpack_scales_q45_K(scales, ksc);
+
+        x_sc[i*(WARP_SIZE/8) + i/8 + ksc] = scales8;
+    }
+#endif // INT8_MMA_AVAILABLE
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q5_K_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const half2 * x_dm = (const half2 *) x_qs + txs.qs;
+    const int   * x_sc = (const int   *) x_dm + txs.dm;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_ds = (const half2 *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR5_K*VDR_Q5_K_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k00/32]) + 2*(k01/16);
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q5_K_q8_1_impl_mmq(
+                    &x_qs[i*(QR5_K*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], sc, sc+8,
+                    x_dm[i], &y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q6_K(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+    int   * x_sc = (int   *) (x_df + WARP_SIZE/QI6_K);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+    int   * x_sc = (int   *) (x_df + txs.dm);
+#endif // INT8_MMA_AVAILABLE
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride;
+
+        const int ql = get_int_b2(bxi->ql, threadIdx.x);
+        const int ql0 = (ql >> 0) & 0x0F0F0F0F;
+        const int ql1 = (ql >> 4) & 0x0F0F0F0F;
+
+        const int qh = get_int_b2(bxi->qh, (QI6_K/4) * (threadIdx.x / (QI6_K/2)) + threadIdx.x % (QI6_K/4));
+        const int qh0 = ((qh >> ((threadIdx.x & 0x08) >> 2)) << 4) & 0x30303030;
+        const int qh1 =  (qh >> ((threadIdx.x & 0x08) >> 2))       & 0x30303030;
+
+        const int kq0 = 2*threadIdx.x - threadIdx.x % (QI6_K/2) + 0;
+        const int kq1 = 2*threadIdx.x - threadIdx.x % (QI6_K/2) + QI6_K/2;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq0] = __vsubss4(ql0 | qh0, 0x20202020);
+        x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq1] = __vsubss4(ql1 | qh1, 0x20202020);
+#else
+        x_qs[i*(2*WARP_SIZE + 1)     + kq0] = __vsubss4(ql0 | qh0, 0x20202020);
+        x_qs[i*(2*WARP_SIZE + 1)     + kq1] = __vsubss4(ql1 | qh1, 0x20202020);
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI6_K;  // == 1 if QK_K == 256
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row; // == 0 if QK_K == 256
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI6_K) {
+        int i = (i0 + threadIdx.y * QI6_K + threadIdx.x / blocks_per_tile_x_row) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + kbxd;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q6_K       + kbxd] = bxi->d;
+#else
+        x_df[i*(WARP_SIZE/QI6_K) + i/QI6_K + kbxd] = bxi->d;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
+        int i = (i0 + threadIdx.y * 8 + threadIdx.x / (WARP_SIZE/8)) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + (threadIdx.x % (WARP_SIZE/8)) / 4;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + threadIdx.x % (WARP_SIZE/8)] = get_int_b2(bxi->scales, threadIdx.x % (QI6_K/8));
+#else
+        x_sc[i*(WARP_SIZE/8) + i/8   + threadIdx.x % (WARP_SIZE/8)] = get_int_b2(bxi->scales, threadIdx.x % (QI6_K/8));
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + txs.qs;
+    const int   * x_sc = (const int   *) x_df + txs.dm;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR6_K*VDR_Q6_K_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const int8_t * sc = ((const int8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k0/16]);
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q6_K_q8_1_impl_mmq(
+                    &x_qs[i*(QR6_K*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], sc,
+                    x_df[i*(WARP_SIZE/QI6_K) + i/QI6_K], &y_df[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+#ifdef INT8_MMA_AVAILABLE
+
+    typedef mma_int_A_I16K4 mma_A;
+    typedef mma_int_B_J8K4  mma_B;
+    typedef mma_int_C_I16J8 mma_C;
+
+    constexpr int granularity = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = 2 * granularity;
+    constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
+
+    y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
+
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + WARP_SIZE*2;
+    const int   * x_sc = (const int   *) x_df + WARP_SIZE/QI6_K;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+
+    const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I);
+
+    mma_A   A[ntx][8];
+    int   scA[ntx][mma_C::ne/2][8];
+    float  dA[ntx][mma_C::ne/2];
+
+#pragma unroll
+    for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += 8) {
+            const int k0 = k00 + k01;
+
+            A[n][k01/4 + 0].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q6_K + (k0 + 0),        MMQ_MMA_TILE_X_K_Q6_K);
+            A[n][k01/4 + 1].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q6_K + (k0 + mma_A::K), MMQ_MMA_TILE_X_K_Q6_K);
+        }
+
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += 16) {
+            const int k0 = k00 + k01;
+
+#pragma unroll
+            for (int l = 0; l < mma_C::ne/2; ++l) {
+                const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
+
+                const int      sc_packed = x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + k0/16];
+                const int8_t * sc        = (const int8_t *) &sc_packed;
+
+#pragma unroll
+                for (int ksc = 0; ksc < sizeof(int); ++ksc) {
+                    scA[n][l][k01/4 + ksc] = sc[ksc];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int l = 0; l < mma_C::ne/2; ++l) {
+            const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
+
+            dA[n][l] = x_df[i*MMQ_MMA_TILE_X_K_Q6_K];
+        }
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
+        float tmp[ntx][mma_C::ne] = {{0.0f}};
+
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += 8) {
+            mma_B B[2];
+            float dB[mma_C::ne/2];
+
+            B[0].load(y_qs + j0*MMQ_TILE_Y_K + 0        + k01, MMQ_TILE_Y_K);
+            B[1].load(y_qs + j0*MMQ_TILE_Y_K + mma_B::K + k01, MMQ_TILE_Y_K);
+
+#pragma unroll
+            for (int l = 0; l < mma_C::ne/2; ++l) {
+                const int j = j0 + mma_C::get_j(l);
+
+                dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1];
+            }
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+                mma_C C[2];
+                C[0].mma_K4(A[n][k01/4 + 0], B[0]);
+                C[1].mma_K4(A[n][k01/4 + 1], B[1]);
+
+#pragma unroll
+                for (int l = 0; l < mma_C::ne; ++l) {
+                    tmp[n][l] += (C[0].x[l]*scA[n][l/2][k01/4 + 0] + C[1].x[l]*scA[n][l/2][k01/4 + 1])*dB[l%2];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+            for (int l = 0; l < mma_C::ne; ++l) {
+                sum[(j0/mma_C::J + n)*mma_C::ne + l] += tmp[n][l]*dA[n][l/2];
+            }
+        }
+    }
+#else
+    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
+    NO_DEVICE_CODE;
+#endif // INT8_MMA_AVAILABLE
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq4_nl(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_NL, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kbx  = threadIdx.x / QI4_NL;
+    const int kqsx = threadIdx.x % QI4_NL;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbx;
+
+        const int aux_q4 = get_int_b2(bxi->qs, kqsx);
+        const int2 v = get_int_from_table_16(aux_q4);
+        const int k0 = 8 * (threadIdx.x / 4) + threadIdx.x % 4;
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 0] = v.x;
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 4] = v.y;
+#else
+        x_qs[i*(2*WARP_SIZE + 1)     + k0 + 0] = v.x;
+        x_qs[i*(2*WARP_SIZE + 1)     + k0 + 4] = v.y;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI4_NL;
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_NL) {
+        int i = i0 + threadIdx.y * QI4_NL + threadIdx.x / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbxd;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kbxd] = __half2float(bxi->d);
+#else
+        x_df[i*(WARP_SIZE/4) + i/4   + kbxd] = __half2float(bxi->d);
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_xxs(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_XXS, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % (QI2_XXS/2);
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/(QI2_XXS/2)) {
+        int i = i0 + threadIdx.y*(2*WARP_SIZE/QI2_XXS) + threadIdx.x/(QI2_XXS/2);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq2_xxs * bxi = (const block_iq2_xxs *) x + kbx0 + i*stride;
+
+        const int q2 = get_int_b2(bxi->qs, 2*kqsx+0);
+        const uint8_t * aux8 = (const uint8_t *) &q2;
+        const uint32_t aux32 = get_int_b2(bxi->qs, 2*kqsx+1);
+
+#pragma unroll
+        for (int l = 0; l < QR2_XXS; ++l) {
+            const int * grid_pos = (const int *) (iq2xxs_grid + aux8[l]);
+            const int signs_packed = ksigns_iq2xs[(aux32 >> (7*l)) & 0x7F];
+
+            const int signs0 = __vcmpne4(((signs_packed & 0x03) << 7) | ((signs_packed & 0x0C) << 21), 0x00000000);
+            const int grid0 = __vsub4(grid_pos[0] ^ signs0, signs0);
+
+            const int signs1 = __vcmpne4(((signs_packed & 0x30) << 3) | ((signs_packed & 0xC0) << 17), 0x00000000);
+            const int grid1 = __vsub4(grid_pos[1] ^ signs1, signs1);
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid0;
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 1)] = grid1;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 0)] = grid0;
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 1)] = grid1;
+#endif // INT8_MMA_AVAILABLE
+        }
+
+        const int ls = aux32 >> 28;
+        const float d = bxi->d;
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = (ls*d + d/2)/4;
+#else
+        x_df[i*(WARP_SIZE/4) + i/4   + kqsx] = (ls*d + d/2)/4;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_xs(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16;
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % (QI2_XS/2);
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/(QI2_XS/2)) {
+        int i = i0 + threadIdx.y*(2*WARP_SIZE/QI2_XS) + threadIdx.x/(QI2_XS/2);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq2_xs * bxi = (const block_iq2_xs *) x + kbx0 + i*stride;
+
+        const int2 q2_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1));
+        const uint16_t * q2 = (const uint16_t *) &q2_packed;
+
+    #pragma unroll
+        for (int l = 0; l < QR2_XS; ++l) {
+            const uint32_t * grid_pos = (const uint32_t *)(iq2xs_grid + (q2[l] & 0x000001FF));
+            const uint32_t * signs    = (const uint32_t *)(ksigns64   + (q2[l] >> 9));
+
+            const int grid_l = __vsub4(grid_pos[0] ^ signs[0], signs[0]);
+            const int grid_h = __vsub4(grid_pos[1] ^ signs[1], signs[1]);
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l;
+            x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 1)] = grid_h;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 0)] = grid_l;
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 1)] = grid_h;
+#endif // INT8_MMA_AVAILABLE
+        }
+
+        const int ls = bxi->scales[kqsx];
+        const float d = bxi->d;
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q3_K               + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
+        x_df[i*MMQ_MMA_TILE_X_K_Q3_K               + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
+#else
+        x_df[i*(2*WARP_SIZE*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
+        x_df[i*(2*WARP_SIZE*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_s(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_S, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % (QI2_S/2);
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/(QI2_S/2)) {
+        int i = i0 + threadIdx.y*(2*WARP_SIZE/QI2_S) + threadIdx.x/(QI2_S/2);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq2_s * bxi = (const block_iq2_s *) x + kbx0 + i*stride;
+
+        const int       qs_packed = get_int_b2(bxi->qs, kqsx);
+        const uint8_t * qs        = (const uint8_t *) &qs_packed;
+
+        const int qh = bxi->qh[kqsx];
+
+        const int       signs_packed_32 = get_int_b2(bxi->qs, QK_K/32 + kqsx);
+        const uint8_t * signs_packed_8  = (const uint8_t *) &signs_packed_32;
+
+#pragma unroll
+        for (int l = 0; l < QR2_S; ++l) {
+            const int * grid_pos = (const int *)(iq2s_grid + (qs[l] | ((qh << (8-2*l)) & 0x300)));
+
+            const int signs0 = __vcmpne4(((signs_packed_8[l] & 0x03) << 7) | ((signs_packed_8[l] & 0x0C) << 21), 0x00000000);
+            const int signs1 = __vcmpne4(((signs_packed_8[l] & 0x30) << 3) | ((signs_packed_8[l] & 0xC0) << 17), 0x00000000);
+
+            const int grid_l = __vsub4(grid_pos[0] ^ signs0, signs0);
+            const int grid_h = __vsub4(grid_pos[1] ^ signs1, signs1);
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l;
+            x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 1)] = grid_h;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 0)] = grid_l;
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 1)] = grid_h;
+#endif // INT8_MMA_AVAILABLE
+        }
+
+        const int ls = bxi->scales[kqsx];
+        const float d = bxi->d;
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q3_K               + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
+        x_df[i*MMQ_MMA_TILE_X_K_Q3_K               + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
+#else
+        x_df[i*(2*WARP_SIZE*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
+        x_df[i*(2*WARP_SIZE*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq3_xxs(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_XXS, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % (QI3_XXS/2);
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/(QI3_XXS/2)) {
+        int i = i0 + threadIdx.y*(2*WARP_SIZE/QI3_XXS) + threadIdx.x/(QI3_XXS/2);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq3_xxs * bxi = (const block_iq3_xxs *) x + kbx0 + i*stride;
+
+        const int2 q3_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1));
+        const uint8_t * q3 = (const uint8_t *) &q3_packed;
+        const uint32_t aux32 = get_int_b2(bxi->qs, QK_K/16 + kqsx);
+
+#pragma unroll
+        for (int l = 0; l < QR3_XXS; ++l) {
+            const int2 grid_pos = make_int2(iq3xxs_grid[q3[2*l+0]], iq3xxs_grid[q3[2*l+1]]);
+
+            const int * signs = (const int *)(ksigns64 + ((aux32 >> (7*l)) & 0x7F));
+
+            const int grid_l = __vsub4(grid_pos.x ^ signs[0], signs[0]);
+            const int grid_h = __vsub4(grid_pos.y ^ signs[1], signs[1]);
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid_l;
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 1)] = grid_h;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 0)] = grid_l;
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 1)] = grid_h;
+#endif // INT8_MMA_AVAILABLE
+        }
+
+        const int ls = aux32 >> 28;
+        const float d = bxi->d;
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = (ls*d + d/2)/2;
+#else
+        x_df[i*(WARP_SIZE/4) + i/4   + kqsx] = (ls*d + d/2)/2;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq3_s(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % (QI3_S/2);
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/(QI3_S/2)) {
+        int i = i0 + threadIdx.y*(2*WARP_SIZE/QI3_S) + threadIdx.x/(QI3_S/2);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq3_s * bxi = (const block_iq3_s *) x + kbx0 + i*stride;
+
+        const int2      qs_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1));
+        const uint8_t * qs        = (const uint8_t *) &qs_packed;
+
+        const int qh = bxi->qh[kqsx];
+
+        const int       signs_packed_32 = get_int_b2(bxi->signs, kqsx);
+        const uint8_t * signs_packed_8  = (const uint8_t *) &signs_packed_32;
+
+#pragma unroll
+        for (int l = 0; l < QR3_S; ++l) {
+            const int2 grid_pos = make_int2(
+                iq3s_grid[qs[2*l+0] | ((qh << (8 - 2*l)) & 0x100)],
+                iq3s_grid[qs[2*l+1] | ((qh << (7 - 2*l)) & 0x100)]);
+
+            const int signs0 = __vcmpne4(((signs_packed_8[l] & 0x03) << 7) | ((signs_packed_8[l] & 0x0C) << 21), 0x00000000);
+            const int signs1 = __vcmpne4(((signs_packed_8[l] & 0x30) << 3) | ((signs_packed_8[l] & 0xC0) << 17), 0x00000000);
+
+            const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
+            const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l+0)] = grid_l;
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l+1)] = grid_h;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l+0)] = grid_l;
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l+1)] = grid_h;
+#endif // INT8_MMA_AVAILABLE
+        }
+
+        const int ls = 1 + 2*((bxi->scales[kqsx/2] >> (((2*kqsx) << 1) & 0x04)) & 0x0F);
+        const float d = bxi->d;
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = ls*d;
+#else
+        x_df[i*(WARP_SIZE/4) + i/4   + kqsx] = ls*d;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq1_s(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_ds = (half2 *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_ds = (half2 *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % QI1_S;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/QI1_S) {
+        int i = i0 + threadIdx.y*(WARP_SIZE/QI1_S) + threadIdx.x/QI1_S;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq1_s * bxi = (const block_iq1_s *) x + kbx0 + i*stride;
+
+        const int       qs_packed = get_int_b2(bxi->qs, kqsx);
+        const uint8_t * qs        = (const uint8_t *) &qs_packed;
+
+        const int qh = bxi->qh[kqsx];
+
+    #pragma unroll
+        for (int l = 0; l < QR1_S/2; ++l) {
+            const int grid = iq1s_grid_gpu[qs[l] | (((qh >> (3*l)) & 0x07) << 8)];
+
+            const int grid0 = (grid >> 0) & 0x0F0F0F0F;
+            const int grid1 = (grid >> 4) & 0x0F0F0F0F;
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 8*kqsx + (2*l+0)] = grid0;
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 8*kqsx + (2*l+1)] = grid1;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l+0)] = grid0;
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l+1)] = grid1;
+#endif // INT8_MMA_AVAILABLE
+        }
+
+        const float  d1q   = __half2float(bxi->d) * (((qh >> 11) & 0x0E) + 1);
+        const float  delta = -1.0f + IQ1S_DELTA - (qh & 0x8000) * (2.0f*IQ1S_DELTA/0x8000);
+
+#ifdef INT8_MMA_AVAILABLE
+        x_ds[i*MMQ_MMA_TILE_X_K_Q8_1 + kqsx] = make_half2(d1q, d1q*delta);
+#else
+        x_ds[i*(WARP_SIZE/4) + i/4   + kqsx] = make_half2(d1q, d1q*delta);
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq4_xs(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_XS, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kbx  = 0;           // threadIdx.x / QI4_XS
+    const int kqsx = threadIdx.x; // threadIdx.x % QI4_XS
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride + kbx;
+
+        const int aux_q4 = get_int_b4(bxi->qs, kqsx);
+        const int2 v = get_int_from_table_16(aux_q4);
+        const int k0 = 8 * (threadIdx.x / 4) + threadIdx.x % 4;
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 0] = v.x;
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 4] = v.y;
+#else
+        x_qs[i*(2*WARP_SIZE + 1)     + k0 + 0] = v.x;
+        x_qs[i*(2*WARP_SIZE + 1)     + k0 + 4] = v.y;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) {
+        int i = i0 + threadIdx.y * 4 + threadIdx.x / (WARP_SIZE/4);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride;
+
+        const float d = __half2float(bxi->d);
+
+        const int ls = ((bxi->scales_l[(threadIdx.x % 8)/2] >> (4*(threadIdx.x % 2))) & 0x0F)
+            | (((bxi->scales_h >> (2*(threadIdx.x % 8))) & 0x03) << 4);
+
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + threadIdx.x % 8] = d * (ls - 32);
+#else
+        x_df[i*(WARP_SIZE/4) + i/4   + threadIdx.x % 8] = d * (ls - 32);
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template<int mmq_x, int mmq_y, int nwarps, bool need_check>
+static __device__ __forceinline__ void mmq_write_back_dp4a(
+    const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) {
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+        if (j > j_max) {
+            return;
+        }
+
+#pragma unroll
+        for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            if (need_check && i > i_max) {
+                continue;
+            }
+
+            dst[j*stride + i] = sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
+        }
+    }
+}
+
+template<int mmq_x, int mmq_y, int nwarps, bool need_check>
+static __device__ __forceinline__ void mmq_write_back_mma(
+    const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) {
+
+    typedef mma_int_C_I16J8 mma_C;
+
+    constexpr int granularity = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = 2 * granularity;
+    constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
+
+    const int i0 = (threadIdx.y / ntx) * (ntx*mma_C::I);
+#ifdef INT8_MMA_AVAILABLE
+    static_assert(nwarps*mma_C::I == mmq_y, "nwarps*mma_C::I != mmq_y");
+#endif // INT8_MMA_AVAILABLE
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
+#pragma unroll
+        for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+            for (int l = 0; l < mma_C::ne; ++l) {
+                const int j = j0 + (threadIdx.y % ntx) * mma_C::J + mma_C::get_j(l);
+
+                if (j > j_max) {
+                    continue;
+                }
+
+                const int i = i0 + n*mma_C::I + mma_C::get_i(l);
+
+                if (need_check && i > i_max) {
+                    continue;
+                }
+
+                dst[j*stride + i] = sum[(j0/mma_C::J + n)*mma_C::ne + l];
+            }
+        }
+    }
+}
+
+// -------------------------------------------------------------------------------------------------------------------------------------
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check, ggml_type type>
+struct mmq_type_traits;
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q4_0> {
+    static constexpr int              vdr          = VDR_Q4_0_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q4_0<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_DS4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q4_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q4_1> {
+    static constexpr int              vdr          = VDR_Q4_1_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q4_1<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_1_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q4_1_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q5_0> {
+    static constexpr int              vdr          = VDR_Q5_0_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q5_0<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q5_1> {
+    static constexpr int              vdr          = VDR_Q5_1_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q5_1<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_1_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_1_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q8_0> {
+    static constexpr int              vdr          = VDR_Q8_0_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q8_0<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q2_K> {
+    static constexpr int              vdr          = VDR_Q2_K_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q2_K<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q2_K_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q2_K_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q3_K> {
+    static constexpr int              vdr          = VDR_Q3_K_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q3_K<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_16_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q3_K_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q4_K> {
+    static constexpr int              vdr          = VDR_Q4_K_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q4_K<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_1_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q4_K_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q5_K> {
+    static constexpr int              vdr          = VDR_Q5_K_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q5_K<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_1_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q5_K_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q6_K> {
+    static constexpr int              vdr          = VDR_Q6_K_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q6_K<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q6_K_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q6_K_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ2_XXS> {
+    static constexpr int              vdr          = VDR_IQ2_XXS_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq2_xxs<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ2_XS> {
+    static constexpr int              vdr          = VDR_IQ2_XS_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq2_xs<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_16_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_16_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ2_S> {
+    static constexpr int              vdr          = VDR_IQ2_S_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq2_s<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_16_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_16_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ3_XXS> {
+    static constexpr int              vdr          = VDR_IQ3_XXS_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq3_xxs<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ3_S> {
+    static constexpr int              vdr          = VDR_IQ3_S_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq3_s<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ1_S> {
+    static constexpr int              vdr          = VDR_IQ1_S_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq1_s<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_1_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_1_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ4_NL> {
+    static constexpr int              vdr          = VDR_IQ4_NL_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq4_nl<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ4_XS> {
+    static constexpr int              vdr          = VDR_IQ4_XS_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq4_xs<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <ggml_type type, int mmq_x, int nwarps, bool need_check, bool fixup>
+static __device__ void mul_mat_q_process_tile(
+    const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
+    const int & ne00, const int & ne01, const int & stride01, const int & ne10, const int & ne11, const int & stride11, const int & ne0,
+    const int & it, const int & jt, const int & kb0_start, const int & kb0_stop) {
+
+    constexpr int              qk         = ggml_cuda_type_traits<type>::qk;
+    constexpr int              mmq_y      = get_mmq_y_device();
+    constexpr load_tiles_mmq_t load_tiles = mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, type>::load_tiles;
+
+    extern __shared__ char data_mul_mat_q[];
+    int * tile_y = (int *) data_mul_mat_q;
+    int * tile_x = tile_y + GGML_PAD(mmq_x*(WARP_SIZE + WARP_SIZE/QI8_1), nwarps*WARP_SIZE);
+
+#ifdef INT8_MMA_AVAILABLE
+    constexpr vec_dot_mmq_t    vec_dot    = mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, type>::vec_dot_mma;
+    constexpr mmq_write_back_t write_back = mmq_write_back_mma<mmq_x, mmq_y, nwarps, need_check>;
+#else
+    constexpr vec_dot_mmq_t    vec_dot    = mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, type>::vec_dot_dp4a;
+    constexpr mmq_write_back_t write_back = mmq_write_back_dp4a<mmq_x, mmq_y, nwarps, need_check>;
+#endif // INT8_MMA_AVAILABLE
+
+    constexpr int blocks_per_iter = MMQ_ITER_K / qk;
+
+    float sum[mmq_x*mmq_y / (nwarps*WARP_SIZE)] = {0.0f};
+
+    const int tile_x_max_i = ne01 - it*mmq_y - 1;
+    const int tile_y_max_j = ne11 - jt*mmq_x - 1;
+
+    const int * y = (const int *) yc + jt*(mmq_x*sizeof(block_q8_1_mmq)/sizeof(int));
+
+    for (int kb0 = kb0_start; kb0 < kb0_stop; kb0 += blocks_per_iter) {
+        load_tiles(x, tile_x, stride01*it*mmq_y + kb0, tile_x_max_i, stride01);
+
+        {
+            const int * by0 = y + stride11*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 0*sizeof(block_q8_1_mmq)/sizeof(int));
+#pragma unroll
+            for (int l0 = 0; l0 < mmq_x*MMQ_TILE_Y_K; l0 += nwarps*WARP_SIZE) {
+                int l = l0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+                tile_y[l] = by0[l];
+            }
+        }
+
+        __syncthreads();
+
+        vec_dot(tile_x, tile_y, sum, 0);
+
+        __syncthreads();
+
+        {
+            const int * by0 = y + stride11*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 1*sizeof(block_q8_1_mmq)/sizeof(int));
+#pragma unroll
+            for (int l0 = 0; l0 < mmq_x*MMQ_TILE_Y_K; l0 += nwarps*WARP_SIZE) {
+                int l = l0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+                tile_y[l] = by0[l];
+            }
+        }
+
+        __syncthreads();
+
+        vec_dot(tile_x, tile_y, sum, WARP_SIZE);
+
+        __syncthreads();
+    }
+
+    if (fixup) {
+        write_back(sum, tmp_fixup + blockIdx.x*(mmq_x*mmq_y), mmq_y, mmq_y, mmq_x);
+    } else {
+        write_back(sum, dst + jt*mmq_x*ne0 + it*mmq_y, ne0, tile_x_max_i, tile_y_max_j);
+    }
+}
+
+
+// The mul_mat_q kernel implements "stream-k" work partitioning as described in https://arxiv.org/abs/2301.03598
+
+template <ggml_type type, int mmq_x, int nwarps, bool need_check>
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*nwarps, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#else
+#if __CUDA_ARCH__ >= CC_VOLTA
+    __launch_bounds__(WARP_SIZE*nwarps, 1)
+#else
+    __launch_bounds__(WARP_SIZE*nwarps, 2)
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+static __global__ void mul_mat_q(
+    const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
+    const int ne00, const int ne01, const int stride01, const int ne10, const int ne11, const int stride11, const int ne0) {
+
+    // Skip unused template specializations for faster compilation:
+    if (mmq_x > get_mmq_x_max_device() || mmq_x % mmq_get_granularity_device(mmq_x) != 0) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    constexpr int qk    = ggml_cuda_type_traits<type>::qk;
+    constexpr int mmq_y = get_mmq_y_device();
+
+    // On AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead:
+#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
+    {
+        constexpr bool fixup = false;
+        mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
+            (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
+                blockIdx.x, blockIdx.y, 0, ne00/qk);
+        return;
+    }
+#endif // (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
+
+    const     int64_t blocks_per_ne00 = ne00 / qk;
+    constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
+
+    const int ntx = (ne11 + mmq_x - 1) / mmq_x; // Number of tiles x
+    const int nty = (ne01 + mmq_y - 1) / mmq_y; // Number of tiles y
+
+    // kbc == k block continuous, current index in continuous ijk space.
+    int64_t kbc      = (int64_t) blockIdx.x     *blocks_per_ne00*ntx*nty / gridDim.x;
+    int64_t kbc_stop = (int64_t)(blockIdx.x + 1)*blocks_per_ne00*ntx*nty / gridDim.x;
+
+    kbc      -= (kbc      % blocks_per_ne00) % blocks_per_iter;
+    kbc_stop -= (kbc_stop % blocks_per_ne00) % blocks_per_iter;
+
+    // kb0 == k index when doing the matrix multiplication for an output tile.
+    int kb0_start = kbc % blocks_per_ne00;
+    int kb0_stop  = min(blocks_per_ne00, kb0_start + kbc_stop - kbc);
+    while (kbc < kbc_stop && kb0_stop == blocks_per_ne00) {
+        const int jt =  kbc /    (blocks_per_ne00*nty);                    // j index of current tile.
+        const int it = (kbc - jt*(blocks_per_ne00*nty)) / blocks_per_ne00; // i index of current tile.
+
+        constexpr bool fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
+        mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
+            (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
+             it, jt, kb0_start, kb0_stop);
+
+        kbc += blocks_per_ne00;
+        kbc -= kbc % blocks_per_ne00;
+
+        kb0_start = 0;
+        kb0_stop  = min(blocks_per_ne00, kbc_stop - kbc);
+    }
+
+    if (kbc >= kbc_stop) {
+        return;
+    }
+
+    const int jt =  kbc /    (blocks_per_ne00*nty);
+    const int it = (kbc - jt*(blocks_per_ne00*nty)) / blocks_per_ne00;
+
+    constexpr bool fixup = true; // Last index writes it data to fixup buffer to avoid data races with other blocks.
+    mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
+        (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
+            it, jt, kb0_start, kb0_stop);
+}
+
+
+template <ggml_type type, int mmq_x, int nwarps, bool need_check>
+static __global__ void mul_mat_q_stream_k_fixup(
+    float * __restrict__ dst, const float * __restrict__ tmp_last_tile, const int ne00, const int ne01, const int ne11, const int ne0, const int block_num_mmq) {
+
+    constexpr int     mmq_y           = get_mmq_y_device();
+    constexpr int     qk              = ggml_cuda_type_traits<type>::qk;
+    constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
+    const     int64_t blocks_per_ne00 = ne00 / qk;
+
+    float sum[mmq_x*mmq_y / (nwarps*WARP_SIZE)] = {0.0f};
+
+    const int ntx = (ne11 + mmq_x - 1) / mmq_x;
+    const int nty = (ne01 + mmq_y - 1) / mmq_y;
+
+    bool any_fixup = false;
+
+    const int bidx_start = ((blockIdx.y*nty + blockIdx.x)     * block_num_mmq)                           / (gridDim.y*gridDim.x);
+    const int bidx_stop  = ((blockIdx.y*nty + blockIdx.x + 1) * block_num_mmq + gridDim.y*gridDim.x - 1) / (gridDim.y*gridDim.x);
+
+    int64_t kbc_0;
+    int64_t kbc_stop_0 = (int64_t) bidx_start*blocks_per_ne00*ntx*nty / block_num_mmq;
+
+    for (int bidx = bidx_start; bidx < bidx_stop; ++bidx) {
+        kbc_0 = kbc_stop_0;
+        kbc_stop_0 = (int64_t) (bidx + 1)*blocks_per_ne00*ntx*nty / block_num_mmq;
+
+        const int64_t kbc      = kbc_0      - (kbc_0      % blocks_per_ne00) % blocks_per_iter;
+        const int64_t kbc_stop = kbc_stop_0 - (kbc_stop_0 % blocks_per_ne00) % blocks_per_iter;
+
+        // Skip fixup tile if the MMQ CUDA block never wrote anything to it:
+        if (kbc == kbc_stop || kbc_stop % blocks_per_ne00 == 0) {
+            continue;
+        }
+
+        const int jt =  kbc_stop /    (blocks_per_ne00*nty);
+        const int it = (kbc_stop - jt*(blocks_per_ne00*nty)) / blocks_per_ne00;
+
+        // Skip fixup tile if it's unrelated to the output tile assigned to this CUDA block:
+        if (it != blockIdx.x || jt != blockIdx.y) {
+            continue;
+        }
+
+        any_fixup = true;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE] += tmp_last_tile[bidx*(mmq_x*mmq_y) + j*mmq_y + i];
+            }
+        }
+    }
+
+    if (!any_fixup) {
+        return;
+    }
+
+    dst += blockIdx.y*mmq_x*ne0 + blockIdx.x*mmq_y;
+
+    const int i_max = ne01 - blockIdx.x*mmq_y - 1;
+    const int j_max = ne11 - blockIdx.y*mmq_x - 1;
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+        if (j > j_max) {
+            return;
+        }
+
+#pragma unroll
+        for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            if (need_check && i > i_max) {
+                continue;
+            }
+
+            dst[j*ne0 + i] += sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
+        }
+    }
+}
+
+struct mmq_args {
+    const char * x; const char * y; float * dst;
+    int64_t ne00; int64_t ne01; int64_t stride01;
+    int64_t ne10; int64_t ne11; int64_t stride11;
+    int64_t ne0;
+};
+
+template<ggml_type type>
+static int mmq_get_shmem(const int mmq_x, const int mmq_y, const int cc) {
+    const tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(type, mmq_y);
+    const int mmq_tile_x_k = mmq_get_mma_tile_x_k(type);
+    const int shmem_x = int8_mma_available(cc) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
+    const int shmem_y = mmq_x*sizeof(block_q8_1_mmq);
+    return shmem_x + GGML_PAD(shmem_y, MMQ_NWARPS*WARP_SIZE*sizeof(int));
+}
+
+template <ggml_type type, int mmq_x>
+static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
+    const int id = ggml_cuda_get_device();
+    const int cc = ggml_cuda_info().devices[id].cc;
+    const int nsm = ggml_cuda_info().devices[id].nsm;
+    const int mmq_y = get_mmq_y_host(cc);
+
+    const dim3 block_dims(WARP_SIZE, MMQ_NWARPS, 1);
+
+    const int shmem = mmq_get_shmem<type>(mmq_x, mmq_y, cc);
+
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+    static bool shmem_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
+    if (!shmem_limit_raised[id]) {
+        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
+        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
+        shmem_limit_raised[id] = true;
+    }
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+
+    const int nty = (args.ne01 + mmq_y - 1) / mmq_y;
+    const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;
+    const dim3 block_nums_xy_tiling(nty, ntx, 1);
+
+    const bool use_stream_k = cc >= CC_VOLTA && cc < CC_OFFSET_AMD;
+    if (!use_stream_k) {
+        if (args.ne01 % mmq_y == 0) {
+            constexpr bool need_check = false;
+            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>
+                (args.x, args.y, args.dst, nullptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+        } else {
+            constexpr bool need_check = true;
+            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>
+                (args.x, args.y, args.dst, nullptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+        }
+        return;
+    }
+
+    const dim3 block_nums_mmq(nsm, 1, 1);
+
+    ggml_cuda_pool & pool = ctx.pool(id);
+    ggml_cuda_pool_alloc<float> tmp_fixup(pool, block_nums_mmq.x * mmq_x*mmq_y);
+
+    if (args.ne01 % mmq_y == 0) {
+        constexpr bool need_check = false;
+
+        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_mmq, block_dims, shmem, stream>>>
+            (args.x, args.y, args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+
+        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, 0, stream>>>
+            (args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.ne11, args.ne0, block_nums_mmq.x);
+    } else {
+        constexpr bool need_check = true;
+
+        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_mmq, block_dims, shmem, stream>>>
+            (args.x, args.y, args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+
+        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, 0, stream>>>
+            (args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.ne11, args.ne0, block_nums_mmq.x);
+    }
+}
+
+template <ggml_type type>
+void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
+    const int id    = ggml_cuda_get_device();
+    const int nsm   = ggml_cuda_info().devices[id].nsm;
+    const int cc    = ggml_cuda_info().devices[id].cc;
+    const int smpbo = ggml_cuda_info().devices[id].smpbo;
+
+    const int mmq_x_max = get_mmq_x_max_host(cc);
+    const int mmq_y = get_mmq_y_host(cc);
+    const int block_num_y = (args.ne01 + mmq_y - 1) / mmq_y;
+    const bool use_stream_k = cc >= CC_VOLTA && cc < CC_OFFSET_AMD;
+
+    int mmq_x_best  = 0;
+    int nparts_best = INT_MAX;
+
+    for (int mmq_x = 8; mmq_x <= mmq_x_max && nparts_best > 1; mmq_x += 8) {
+        const int granularity = mmq_get_granularity_host(mmq_x, cc);
+
+        if (mmq_x % granularity != 0 || mmq_get_shmem<type>(mmq_x, mmq_y, cc) > smpbo) {
+            continue;
+        }
+
+        const int ntiles_x = (args.ne11 + mmq_x - 1) / mmq_x;
+        const int nwaves_xy_tiling = ntiles_x*block_num_y;
+        const int nparts = use_stream_k ? ntiles_x : nwaves_xy_tiling;
+
+        if (nparts < nparts_best) {
+            mmq_x_best  = mmq_x;
+            nparts_best = nparts;
+        }
+    }
+
+    switch (mmq_x_best) {
+        case   8:
+            launch_mul_mat_q<type,   8>(ctx, args, stream);
+            break;
+        case  16:
+            launch_mul_mat_q<type,  16>(ctx, args, stream);
+            break;
+        case  24:
+            launch_mul_mat_q<type,  24>(ctx, args, stream);
+            break;
+        case  32:
+            launch_mul_mat_q<type,  32>(ctx, args, stream);
+            break;
+        case  40:
+            launch_mul_mat_q<type,  40>(ctx, args, stream);
+            break;
+        case  48:
+            launch_mul_mat_q<type,  48>(ctx, args, stream);
+            break;
+        case  56:
+            launch_mul_mat_q<type,  56>(ctx, args, stream);
+            break;
+        case  64:
+            launch_mul_mat_q<type,  64>(ctx, args, stream);
+            break;
+        case  72:
+            launch_mul_mat_q<type,  72>(ctx, args, stream);
+            break;
+        case  80:
+            launch_mul_mat_q<type,  80>(ctx, args, stream);
+            break;
+        case  88:
+            launch_mul_mat_q<type,  88>(ctx, args, stream);
+            break;
+        case  96:
+            launch_mul_mat_q<type,  96>(ctx, args, stream);
+            break;
+        case 104:
+            launch_mul_mat_q<type, 104>(ctx, args, stream);
+            break;
+        case 112:
+            launch_mul_mat_q<type, 112>(ctx, args, stream);
+            break;
+        case 120:
+            launch_mul_mat_q<type, 120>(ctx, args, stream);
+            break;
+        case 128:
+            launch_mul_mat_q<type, 128>(ctx, args, stream);
+            break;
+        default:
+            fprintf(stderr, "mmq_x_best=%d\n", mmq_x_best);
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
+#define DECL_MMQ_CASE(type)                                                        \
+    template void mul_mat_q_case<type>(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) \
+
+extern DECL_MMQ_CASE(GGML_TYPE_Q4_0);
+extern DECL_MMQ_CASE(GGML_TYPE_Q4_1);
+extern DECL_MMQ_CASE(GGML_TYPE_Q5_0);
+extern DECL_MMQ_CASE(GGML_TYPE_Q5_1);
+extern DECL_MMQ_CASE(GGML_TYPE_Q8_0);
+extern DECL_MMQ_CASE(GGML_TYPE_Q2_K);
+extern DECL_MMQ_CASE(GGML_TYPE_Q3_K);
+extern DECL_MMQ_CASE(GGML_TYPE_Q4_K);
+extern DECL_MMQ_CASE(GGML_TYPE_Q5_K);
+extern DECL_MMQ_CASE(GGML_TYPE_Q6_K);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ2_XXS);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ2_XS);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ2_S);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ3_XXS);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ3_S);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ1_S);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ4_NL);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ4_XS);
+
+// -------------------------------------------------------------------------------------------------------------------------
+
+void ggml_cuda_op_mul_mat_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream);
+
+bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11);

llama/ggml-cuda/mmvq.cu

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "mmvq.cuh"
+#include "vecdotq.cuh"
+
+typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs);
+
+static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type) {
+    return type == GGML_TYPE_Q4_0 ? vec_dot_q4_0_q8_1 :
+        type == GGML_TYPE_Q4_1 ? vec_dot_q4_1_q8_1 :
+        type == GGML_TYPE_Q5_0 ? vec_dot_q5_0_q8_1 :
+        type == GGML_TYPE_Q5_1 ? vec_dot_q5_1_q8_1 :
+        type == GGML_TYPE_Q8_0 ? vec_dot_q8_0_q8_1 :
+        type == GGML_TYPE_Q2_K ? vec_dot_q2_K_q8_1 :
+        type == GGML_TYPE_Q3_K ? vec_dot_q3_K_q8_1 :
+        type == GGML_TYPE_Q4_K ? vec_dot_q4_K_q8_1 :
+        type == GGML_TYPE_Q5_K ? vec_dot_q5_K_q8_1 :
+        type == GGML_TYPE_Q6_K ? vec_dot_q6_K_q8_1 :
+        type == GGML_TYPE_IQ2_XXS ? vec_dot_iq2_xxs_q8_1 :
+        type == GGML_TYPE_IQ2_XS ? vec_dot_iq2_xs_q8_1 :
+        type == GGML_TYPE_IQ2_S ? vec_dot_iq2_s_q8_1 :
+        type == GGML_TYPE_IQ3_XXS ? vec_dot_iq3_xxs_q8_1 :
+        type == GGML_TYPE_IQ1_S ? vec_dot_iq1_s_q8_1 :
+        type == GGML_TYPE_IQ1_M ? vec_dot_iq1_m_q8_1 :
+        type == GGML_TYPE_IQ4_NL ? vec_dot_iq4_nl_q8_1 :
+        type == GGML_TYPE_IQ4_XS ? vec_dot_iq4_xs_q8_1 :
+        type == GGML_TYPE_IQ3_S ? vec_dot_iq3_s_q8_1 :
+        nullptr;
+}
+
+static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
+    return type == GGML_TYPE_Q4_0 ? VDR_Q4_0_Q8_1_MMVQ :
+        type == GGML_TYPE_Q4_1    ? VDR_Q4_1_Q8_1_MMVQ :
+        type == GGML_TYPE_Q5_0    ? VDR_Q5_0_Q8_1_MMVQ :
+        type == GGML_TYPE_Q5_1    ? VDR_Q5_1_Q8_1_MMVQ :
+        type == GGML_TYPE_Q8_0    ? VDR_Q8_0_Q8_1_MMVQ :
+        type == GGML_TYPE_Q2_K    ? VDR_Q2_K_Q8_1_MMVQ :
+        type == GGML_TYPE_Q3_K    ? VDR_Q3_K_Q8_1_MMVQ :
+        type == GGML_TYPE_Q4_K    ? VDR_Q4_K_Q8_1_MMVQ :
+        type == GGML_TYPE_Q5_K    ? VDR_Q5_K_Q8_1_MMVQ :
+        type == GGML_TYPE_Q6_K    ? VDR_Q6_K_Q8_1_MMVQ :
+        type == GGML_TYPE_IQ2_XXS ? VDR_IQ2_XXS_Q8_1_MMVQ :
+        type == GGML_TYPE_IQ2_XS  ? VDR_IQ2_XS_Q8_1_MMVQ :
+        type == GGML_TYPE_IQ2_S   ? VDR_IQ2_S_Q8_1_MMVQ :
+        type == GGML_TYPE_IQ3_XXS ? VDR_IQ3_XXS_Q8_1_MMVQ :
+        type == GGML_TYPE_IQ3_S   ? VDR_IQ3_S_Q8_1_MMVQ :
+        type == GGML_TYPE_IQ4_NL  ? VDR_IQ4_NL_Q8_1_MMVQ :
+        type == GGML_TYPE_IQ4_XS  ? VDR_IQ4_XS_Q8_1_MMVQ :
+        1;
+}
+
+template <ggml_type type, int ncols_y>
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+// tell the compiler to use as many registers as it wants, see nwarps definition below
+__launch_bounds__((ncols_y <= 4 ? 4 : 2)*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void mul_mat_vec_q(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
+
+    constexpr int qk  = ggml_cuda_type_traits<type>::qk;
+    constexpr int qi  = ggml_cuda_type_traits<type>::qi;
+    constexpr int vdr = get_vdr_mmvq(type);
+
+    constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
+    constexpr int nwarps              = 1;
+    constexpr int rows_per_cuda_block = 1;
+#else
+    constexpr int nwarps              = ncols_y <= 4 ? 4 : 2;
+    constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
+
+    const     int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    const     int row0 = rows_per_cuda_block*blockIdx.x;
+    const     int blocks_per_row_x = ncols_x / qk;
+    const     int blocks_per_col_y = nrows_y / QK8_1;
+    constexpr int blocks_per_iter = vdr * nwarps*WARP_SIZE / qi;
+
+// partial sum for each thread
+    float tmp[ncols_y][rows_per_cuda_block] = {0.0f};
+
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
+        const int kby = kbx * (qk/QK8_1); // y block index that aligns with kbx
+
+        // x block quant index when casting the quants to int
+        const int kqs = vdr * (tid % (qi/vdr));
+
+#pragma unroll
+        for (int j = 0; j < ncols_y; ++j) {
+#pragma unroll
+            for (int i = 0; i < rows_per_cuda_block; ++i) {
+                tmp[j][i] += vec_dot_q_cuda(vx, &y[j*blocks_per_col_y + kby], (row0 + i)*blocks_per_row_x + kbx, kqs);
+            }
+        }
+    }
+
+    __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][WARP_SIZE];
+    if (threadIdx.y > 0) {
+#pragma unroll
+        for (int j = 0; j < ncols_y; ++j) {
+#pragma unroll
+            for (int i = 0; i < rows_per_cuda_block; ++i) {
+                tmp_shared[threadIdx.y-1][j][i][threadIdx.x] = tmp[j][i];
+            }
+        }
+    }
+    __syncthreads();
+    if (threadIdx.y > 0) {
+        return;
+    }
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int j = 0; j < ncols_y; ++j) {
+#pragma unroll
+        for (int i = 0; i < rows_per_cuda_block; ++i) {
+#pragma unroll
+            for (int l = 0; l < nwarps-1; ++l) {
+                tmp[j][i] += tmp_shared[l][j][i][threadIdx.x];
+            }
+            tmp[j][i] = warp_reduce_sum(tmp[j][i]);
+        }
+
+        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < nrows_dst)) {
+            dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x];
+        }
+    }
+}
+
+template <ggml_type type>
+static void mul_mat_vec_q_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    GGML_ASSERT(ncols_x % ggml_blck_size(type) == 0);
+    GGML_ASSERT(ncols_y <= MMVQ_MAX_BATCH_SIZE);
+
+    int id = ggml_cuda_get_device();
+
+    int64_t nwarps = 1;
+    int64_t rows_per_cuda_block = 1;
+
+    if (ggml_cuda_info().devices[id].cc < CC_RDNA2) { // NVIDIA and AMD older than RDNA2
+        switch(ncols_y) {
+            case 1:
+                nwarps = 4;
+                rows_per_cuda_block = 1;
+                break;
+            case 2:
+            case 3:
+            case 4:
+                nwarps = 4;
+                rows_per_cuda_block = 2;
+                break;
+            case 5:
+            case 6:
+            case 7:
+            case 8:
+                nwarps = 2;
+                rows_per_cuda_block = 2;
+                break;
+            default:
+                GGML_ABORT("fatal error");
+                break;
+        }
+    }
+    const int64_t nblocks = (nrows_x + rows_per_cuda_block - 1) / rows_per_cuda_block;
+    const dim3 block_nums(nblocks, 1, 1);
+    const dim3 block_dims(WARP_SIZE, nwarps, 1);
+
+    switch (ncols_y) {
+        case 1:
+            mul_mat_vec_q<type, 1><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 2:
+            mul_mat_vec_q<type, 2><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 3:
+            mul_mat_vec_q<type, 3><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 4:
+            mul_mat_vec_q<type, 4><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 5:
+            mul_mat_vec_q<type, 5><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 6:
+            mul_mat_vec_q<type, 6><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 7:
+            mul_mat_vec_q<type, 7><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 8:
+            mul_mat_vec_q<type, 8><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
+static void mul_mat_vec_q4_0_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q4_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q4_1_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q4_1>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q5_0_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q5_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q5_1_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q5_1>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q8_0_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q8_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q2_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q2_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q3_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q3_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q4_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q4_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q5_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q5_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q6_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q6_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq2_xxs_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ2_XXS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq2_xs_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ2_XS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq2_s_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ2_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq3_xxs_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ3_XXS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq1_s_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ1_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq1_m_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ1_M>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq4_nl_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ4_NL>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq4_xs_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ4_XS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq3_s_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ3_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+void ggml_cuda_op_mul_mat_vec_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t row_diff = row_high - row_low;
+
+    const int64_t ne10 = src1->ne[0];
+    GGML_ASSERT(ne10 % QK8_1 == 0);
+
+    const int64_t ne0 = dst->ne[0];
+
+    int id = ggml_cuda_get_device();
+
+    // the main device has a larger memory buffer to hold the results from all GPUs
+    // nrows_dst == nrows of the matrix that the kernel writes into
+    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
+
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0:
+            mul_mat_vec_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            mul_mat_vec_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            mul_mat_vec_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            mul_mat_vec_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            mul_mat_vec_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            mul_mat_vec_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            mul_mat_vec_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            mul_mat_vec_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            mul_mat_vec_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            mul_mat_vec_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ2_XXS:
+            mul_mat_vec_iq2_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ2_XS:
+            mul_mat_vec_iq2_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ2_S:
+            mul_mat_vec_iq2_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ3_XXS:
+            mul_mat_vec_iq3_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ1_S:
+            mul_mat_vec_iq1_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ1_M:
+            mul_mat_vec_iq1_m_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ4_NL:
+            mul_mat_vec_iq4_nl_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ4_XS:
+            mul_mat_vec_iq4_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ3_S:
+            mul_mat_vec_iq3_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddf_i);
+    GGML_UNUSED(src1_ncols);
+    GGML_UNUSED(src1_padded_row_size);
+}

llama/ggml-cuda/mmvq.cuh

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "common.cuh"
+
+#define MMVQ_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVQ kernels.
+
+void ggml_cuda_op_mul_mat_vec_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream);

llama/ggml-cuda/norm.cu

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "norm.cuh"
+
+template <int block_size>
+static __global__ void norm_f32(const float * x, float * dst, const int ncols, const float eps) {
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    const int tid = threadIdx.x;
+
+    float2 mean_var = make_float2(0.f, 0.f);
+
+    for (int col = tid; col < ncols; col += block_size) {
+        const float xi = x[row*ncols + col];
+        mean_var.x += xi;
+        mean_var.y += xi * xi;
+    }
+
+    // sum up partial sums
+    mean_var = warp_reduce_sum(mean_var);
+    if (block_size > WARP_SIZE) {
+        __shared__ float2 s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = mean_var;
+        }
+        __syncthreads();
+        mean_var = s_sum[lane_id];
+        mean_var = warp_reduce_sum(mean_var);
+    }
+
+    const float mean = mean_var.x / ncols;
+    const float var = mean_var.y / ncols - mean * mean;
+    const float inv_std = rsqrtf(var + eps);
+
+    for (int col = tid; col < ncols; col += block_size) {
+        dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_std;
+    }
+}
+
+template <int block_size>
+static __global__ void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps) {
+    // blockIdx.x: num_groups idx
+    // threadIdx.x: block_size idx
+    int start = blockIdx.x * group_size;
+    int end = start + group_size;
+
+    start += threadIdx.x;
+
+    if (end >= ne_elements) {
+        end = ne_elements;
+    }
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int j = start; j < end; j += block_size) {
+        tmp += x[j];
+    }
+
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __shared__ float s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    float mean = tmp / group_size;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += block_size) {
+        float xi = x[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __shared__ float s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    float variance = tmp / group_size;
+    float scale = rsqrtf(variance + eps);
+    for (int j = start; j < end; j += block_size) {
+        dst[j] *= scale;
+    }
+}
+
+template <int block_size>
+static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps) {
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    const int tid = threadIdx.x;
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int col = tid; col < ncols; col += block_size) {
+        const float xi = x[row*ncols + col];
+        tmp += xi * xi;
+    }
+
+    // sum up partial sums
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __shared__ float s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    const float mean = tmp / ncols;
+    const float scale = rsqrtf(mean + eps);
+
+    for (int col = tid; col < ncols; col += block_size) {
+        dst[row*ncols + col] = scale * x[row*ncols + col];
+    }
+}
+
+static void norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
+    GGML_ASSERT(ncols % WARP_SIZE == 0);
+    if (ncols < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    }
+}
+
+static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const float eps, const int group_size, const int ne_elements, cudaStream_t stream) {
+    if (group_size < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        group_norm_f32<1024><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
+    }
+}
+
+static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
+    GGML_ASSERT(ncols % WARP_SIZE == 0);
+    if (ncols < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        rms_norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        rms_norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    }
+}
+
+void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    norm_f32_cuda(src0_d, dst_d, ne00, nrows, eps, stream);
+}
+
+void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    int num_groups = dst->op_params[0];
+
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
+    int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
+    group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], eps, group_size, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    rms_norm_f32_cuda(src0_d, dst_d, ne00, nrows, eps, stream);
+}

llama/ggml-cuda/norm.cuh

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "common.cuh"
+
+void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/pad.cu

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "pad.cuh"
+
+static __global__ void pad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02, const int ne03) {
+    // blockIdx.z: idx of ne2*ne3, aka ne02*ne03
+    // blockIdx.y: idx of ne1
+    // blockIDx.x: idx of ne0 / BLOCK_SIZE
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+
+    // operation
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) {
+        int offset_src =
+            nidx +
+            blockIdx.y * ne00 +
+            blockIdx.z * ne00 * ne01;
+        dst[offset_dst] = x[offset_src];
+    } else {
+        dst[offset_dst] = 0.0f;
+    }
+}
+
+static void pad_f32_cuda(const float * x, float * dst,
+    const int ne00, const int ne01, const int ne02, const int ne03,
+    const int ne0, const int ne1, const int ne2, const int ne3, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne1, ne2*ne3);
+    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02, ne03);
+}
+
+void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
+
+    pad_f32_cuda(src0_d, dst_d,
+        src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
+}

llama/ggml-cuda/pad.cuh

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "common.cuh"
+
+#define CUDA_PAD_BLOCK_SIZE 256
+
+void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/pool2d.cu

+/**
+ * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "pool2d.cuh"
+
+template <typename Ti, typename To>
+static  __global__ void pool2d_nchw_kernel(
+        const int ih, const int iw, const int oh, const int ow,
+        const int kh, const int kw, const int sh, const int sw,
+        const int ph, const int pw, const int parallel_elements,
+        const Ti* src, To* dst, const enum ggml_op_pool op) {
+    int idx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (idx >= parallel_elements) {
+        return;
+    }
+
+    const int I_HW = ih * iw;
+    const int O_HW = oh * ow;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / ow;
+    const int cur_ow = idx % O_HW % ow;
+    const Ti* i_ptr = src + nc * I_HW;
+    To* o_ptr = dst + nc * O_HW;
+    const int start_h = cur_oh * sh - ph;
+    const int bh = max(0, start_h);
+    const int eh = min(ih, start_h + kh);
+    const int start_w = cur_ow * sw - pw;
+    const int bw = max(0, start_w);
+    const int ew = min(iw, start_w + kw);
+    const To scale = 1. / (kh * kw);
+    To res = 0;
+
+    switch (op) {
+        case GGML_OP_POOL_AVG: res = 0; break;
+        case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
+        default: assert(false);
+    }
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+#if __CUDA_ARCH__ >= 350
+            Ti cur = __ldg(i_ptr + i * iw + j);
+#else
+            Ti cur = i_ptr[i * iw + j];
+#endif
+            switch (op) {
+                case GGML_OP_POOL_AVG: res += cur * scale; break;
+                case GGML_OP_POOL_MAX: res = max(res, (To)cur); break;
+                default: assert(false);
+            }
+        }
+    }
+    o_ptr[cur_oh * ow + cur_ow] = res;
+}
+
+static void pool2d_nchw_kernel_f32_f32_cuda(
+        const int ih, const int iw, const int oh, const int ow,
+        const int kh, const int kw, const int sh, const int sw,
+        const int ph, const int pw, const int parallel_elements,
+        const float * src, float * dst, const enum ggml_op_pool op,
+        cudaStream_t stream) {
+
+    const int num_blocks = (parallel_elements + CUDA_POOL2D_BLOCK_SIZE - 1) / CUDA_POOL2D_BLOCK_SIZE;
+    dim3 block_nums(num_blocks);
+    pool2d_nchw_kernel<<<block_nums, CUDA_POOL2D_BLOCK_SIZE, 0, stream>>>(ih, iw, oh, ow, kh, kw, sh, sw, ph, pw, parallel_elements, src, dst, op);
+}
+
+void ggml_cuda_op_pool2d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int32_t * opts = (const int32_t *)dst->op_params;
+    enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
+    const int k0 = opts[1];
+    const int k1 = opts[2];
+    const int s0 = opts[3];
+    const int s1 = opts[4];
+    const int p0 = opts[5];
+    const int p1 = opts[6];
+
+    const int64_t IH = src0->ne[1];
+    const int64_t IW = src0->ne[0];
+
+    const int64_t N = dst->ne[3];
+    const int64_t OC = dst->ne[2];
+    const int64_t OH = dst->ne[1];
+    const int64_t OW = dst->ne[0];
+
+    const int parallel_elements = N * OC * OH * OW;
+
+    pool2d_nchw_kernel_f32_f32_cuda(IH, IW, OH, OW, k1, k0, s1, s0, p1, p0, parallel_elements, src0_d, dst_d, op, stream);
+}